diff --git a/configs/example_training/autoencoder/kl-f4/imagenet-attnfree-logvar.yaml b/configs/example_training/autoencoder/kl-f4/imagenet-attnfree-logvar.yaml new file mode 100644 index 0000000..731f559 --- /dev/null +++ b/configs/example_training/autoencoder/kl-f4/imagenet-attnfree-logvar.yaml @@ -0,0 +1,104 @@ +model: + base_learning_rate: 4.5e-6 + target: sgm.models.autoencoder.AutoencodingEngine + params: + input_key: jpg + monitor: val/rec_loss + + loss_config: + target: sgm.modules.autoencoding.losses.GeneralLPIPSWithDiscriminator + params: + perceptual_weight: 0.25 + disc_start: 20001 + disc_weight: 0.5 + learn_logvar: True + + regularization_weights: + kl_loss: 1.0 + + regularizer_config: + target: sgm.modules.autoencoding.regularizers.DiagonalGaussianRegularizer + + encoder_config: + target: sgm.modules.diffusionmodules.model.Encoder + params: + attn_type: none + double_z: True + z_channels: 4 + resolution: 256 + in_channels: 3 + out_ch: 3 + ch: 128 + ch_mult: [1, 2, 4] + num_res_blocks: 4 + attn_resolutions: [] + dropout: 0.0 + + decoder_config: + target: sgm.modules.diffusionmodules.model.Decoder + params: ${model.params.encoder_config.params} + +data: + target: sgm.data.dataset.StableDataModuleFromConfig + params: + train: + datapipeline: + urls: + - DATA-PATH + pipeline_config: + shardshuffle: 10000 + sample_shuffle: 10000 + + decoders: + - pil + + postprocessors: + - target: sdata.mappers.TorchVisionImageTransforms + params: + key: jpg + transforms: + - target: torchvision.transforms.Resize + params: + size: 256 + interpolation: 3 + - target: torchvision.transforms.ToTensor + - target: sdata.mappers.Rescaler + - target: sdata.mappers.AddOriginalImageSizeAsTupleAndCropToSquare + params: + h_key: height + w_key: width + + loader: + batch_size: 8 + num_workers: 4 + + +lightning: + strategy: + target: pytorch_lightning.strategies.DDPStrategy + params: + find_unused_parameters: True + + modelcheckpoint: + params: + every_n_train_steps: 5000 + + callbacks: + metrics_over_trainsteps_checkpoint: + params: + every_n_train_steps: 50000 + + image_logger: + target: main.ImageLogger + params: + enable_autocast: False + batch_frequency: 1000 + max_images: 8 + increase_log_steps: True + + trainer: + devices: 0, + limit_val_batches: 50 + benchmark: True + accumulate_grad_batches: 1 + val_check_interval: 10000 \ No newline at end of file diff --git a/configs/example_training/autoencoder/kl-f4/imagenet-kl_f8_8chn.yaml b/configs/example_training/autoencoder/kl-f4/imagenet-kl_f8_8chn.yaml new file mode 100644 index 0000000..39c7c9d --- /dev/null +++ b/configs/example_training/autoencoder/kl-f4/imagenet-kl_f8_8chn.yaml @@ -0,0 +1,105 @@ +model: + base_learning_rate: 4.5e-6 + target: sgm.models.autoencoder.AutoencodingEngine + params: + input_key: jpg + monitor: val/loss/rec + disc_start_iter: 0 + + encoder_config: + target: sgm.modules.diffusionmodules.model.Encoder + params: + attn_type: vanilla-xformers + double_z: true + z_channels: 8 + resolution: 256 + in_channels: 3 + out_ch: 3 + ch: 128 + ch_mult: [1, 2, 4, 4] + num_res_blocks: 2 + attn_resolutions: [] + dropout: 0.0 + + decoder_config: + target: sgm.modules.diffusionmodules.model.Decoder + params: ${model.params.encoder_config.params} + + regularizer_config: + target: sgm.modules.autoencoding.regularizers.DiagonalGaussianRegularizer + + loss_config: + target: sgm.modules.autoencoding.losses.GeneralLPIPSWithDiscriminator + params: + perceptual_weight: 0.25 + disc_start: 20001 + disc_weight: 0.5 + learn_logvar: True + + regularization_weights: + kl_loss: 1.0 + +data: + target: sgm.data.dataset.StableDataModuleFromConfig + params: + train: + datapipeline: + urls: + - DATA-PATH + pipeline_config: + shardshuffle: 10000 + sample_shuffle: 10000 + + decoders: + - pil + + postprocessors: + - target: sdata.mappers.TorchVisionImageTransforms + params: + key: jpg + transforms: + - target: torchvision.transforms.Resize + params: + size: 256 + interpolation: 3 + - target: torchvision.transforms.ToTensor + - target: sdata.mappers.Rescaler + - target: sdata.mappers.AddOriginalImageSizeAsTupleAndCropToSquare + params: + h_key: height + w_key: width + + loader: + batch_size: 8 + num_workers: 4 + + +lightning: + strategy: + target: pytorch_lightning.strategies.DDPStrategy + params: + find_unused_parameters: True + + modelcheckpoint: + params: + every_n_train_steps: 5000 + + callbacks: + metrics_over_trainsteps_checkpoint: + params: + every_n_train_steps: 50000 + + image_logger: + target: main.ImageLogger + params: + enable_autocast: False + batch_frequency: 1000 + max_images: 8 + increase_log_steps: True + + trainer: + devices: 0, + limit_val_batches: 50 + benchmark: True + accumulate_grad_batches: 1 + val_check_interval: 10000 diff --git a/configs/example_training/imagenet-f8_cond.yaml b/configs/example_training/imagenet-f8_cond.yaml new file mode 100644 index 0000000..4f18150 --- /dev/null +++ b/configs/example_training/imagenet-f8_cond.yaml @@ -0,0 +1,185 @@ +model: + base_learning_rate: 1.0e-4 + target: sgm.models.diffusion.DiffusionEngine + params: + scale_factor: 0.13025 + disable_first_stage_autocast: True + log_keys: + - cls + + scheduler_config: + target: sgm.lr_scheduler.LambdaLinearScheduler + params: + warm_up_steps: [10000] + cycle_lengths: [10000000000000] + f_start: [1.e-6] + f_max: [1.] + f_min: [1.] + + denoiser_config: + target: sgm.modules.diffusionmodules.denoiser.DiscreteDenoiser + params: + num_idx: 1000 + + scaling_config: + target: sgm.modules.diffusionmodules.denoiser_scaling.EpsScaling + discretization_config: + target: sgm.modules.diffusionmodules.discretizer.LegacyDDPMDiscretization + + network_config: + target: sgm.modules.diffusionmodules.openaimodel.UNetModel + params: + use_checkpoint: True + in_channels: 4 + out_channels: 4 + model_channels: 256 + attention_resolutions: [1, 2, 4] + num_res_blocks: 2 + channel_mult: [1, 2, 4] + num_head_channels: 64 + num_classes: sequential + adm_in_channels: 1024 + transformer_depth: 1 + context_dim: 1024 + spatial_transformer_attn_type: softmax-xformers + + conditioner_config: + target: sgm.modules.GeneralConditioner + params: + emb_models: + - is_trainable: True + input_key: cls + ucg_rate: 0.2 + target: sgm.modules.encoders.modules.ClassEmbedder + params: + add_sequence_dim: True + embed_dim: 1024 + n_classes: 1000 + + - is_trainable: False + ucg_rate: 0.2 + input_key: original_size_as_tuple + target: sgm.modules.encoders.modules.ConcatTimestepEmbedderND + params: + outdim: 256 + + - is_trainable: False + input_key: crop_coords_top_left + ucg_rate: 0.2 + target: sgm.modules.encoders.modules.ConcatTimestepEmbedderND + params: + outdim: 256 + + first_stage_config: + target: sgm.models.autoencoder.AutoencoderKL + params: + ckpt_path: + embed_dim: 4 + monitor: val/rec_loss + ddconfig: + attn_type: vanilla-xformers + double_z: true + z_channels: 4 + resolution: 256 + in_channels: 3 + out_ch: 3 + ch: 128 + ch_mult: [1, 2, 4, 4] + num_res_blocks: 2 + attn_resolutions: [] + dropout: 0.0 + lossconfig: + target: torch.nn.Identity + + loss_fn_config: + target: sgm.modules.diffusionmodules.loss.StandardDiffusionLoss + params: + loss_weighting_config: + target: sgm.modules.diffusionmodules.loss_weighting.EpsWeighting + sigma_sampler_config: + target: sgm.modules.diffusionmodules.sigma_sampling.DiscreteSampling + params: + num_idx: 1000 + + discretization_config: + target: sgm.modules.diffusionmodules.discretizer.LegacyDDPMDiscretization + + sampler_config: + target: sgm.modules.diffusionmodules.sampling.EulerEDMSampler + params: + num_steps: 50 + + discretization_config: + target: sgm.modules.diffusionmodules.discretizer.LegacyDDPMDiscretization + + guider_config: + target: sgm.modules.diffusionmodules.guiders.VanillaCFG + params: + scale: 5.0 + +data: + target: sgm.data.dataset.StableDataModuleFromConfig + params: + train: + datapipeline: + urls: + # USER: adapt this path the root of your custom dataset + - DATA_PATH + pipeline_config: + shardshuffle: 10000 + sample_shuffle: 10000 # USER: you might wanna adapt depending on your available RAM + + decoders: + - pil + + postprocessors: + - target: sdata.mappers.TorchVisionImageTransforms + params: + key: jpg # USER: you might wanna adapt this for your custom dataset + transforms: + - target: torchvision.transforms.Resize + params: + size: 256 + interpolation: 3 + - target: torchvision.transforms.ToTensor + - target: sdata.mappers.Rescaler + + - target: sdata.mappers.AddOriginalImageSizeAsTupleAndCropToSquare + params: + h_key: height # USER: you might wanna adapt this for your custom dataset + w_key: width # USER: you might wanna adapt this for your custom dataset + + loader: + batch_size: 64 + num_workers: 6 + +lightning: + modelcheckpoint: + params: + every_n_train_steps: 5000 + + callbacks: + metrics_over_trainsteps_checkpoint: + params: + every_n_train_steps: 25000 + + image_logger: + target: main.ImageLogger + params: + disabled: False + enable_autocast: False + batch_frequency: 1000 + max_images: 8 + increase_log_steps: True + log_first_step: False + log_images_kwargs: + use_ema_scope: False + N: 8 + n_rows: 2 + + trainer: + devices: 0, + benchmark: True + num_sanity_val_steps: 0 + accumulate_grad_batches: 1 + max_epochs: 1000 \ No newline at end of file diff --git a/configs/example_training/seva-clipl_dl3dv.yaml b/configs/example_training/seva-clipl_dl3dv.yaml new file mode 100644 index 0000000..aa06fc2 --- /dev/null +++ b/configs/example_training/seva-clipl_dl3dv.yaml @@ -0,0 +1,145 @@ +model: + base_learning_rate: 1.0e-4 + target: sgm.models.diffusion.DiffusionEngine + params: + scale_factor: 0.18215 + disable_first_stage_autocast: True + input_key: clean_latent + network_wrapper: sgm.modules.diffusionmodules.wrappers.SevaWrapper + + + scheduler_config: + target: sgm.lr_scheduler.LambdaLinearScheduler + params: + warm_up_steps: [10000] + cycle_lengths: [10000000000000] + f_start: [1.e-6] + f_max: [1.] + f_min: [1.] + + denoiser_config: + target: sgm.modules.diffusionmodules.denoiser.DiscreteDenoiser + params: + num_idx: 1000 + + scaling_config: + target: sgm.modules.diffusionmodules.denoiser_scaling.EpsScaling + discretization_config: + target: sgm.modules.diffusionmodules.discretizer.LegacyDDPMDiscretization + + network_config: + target: seva.model.Seva + params: + params: + in_channels: 11 + model_channels: 320 + out_channels: 4 + num_frames: 21 + num_res_blocks: 2 + attention_resolutions: [4, 2, 1] + channel_mult: [1, 2, 4, 4] + num_head_channels: 64 + transformer_depth: [1, 1, 1, 1] + context_dim: 1024 + dense_in_channels: 6 + dropout: 0.0 + unflatten_names: ["middle_ds8", "output_ds4", "output_ds2"] + + conditioner_config: + target: sgm.modules.GeneralConditioner + params: + emb_models: + - is_trainable: False + ucg_rate: 0.0 + input_key: plucker + target: sgm.modules.encoders.modules.IdentityEncoder + + - is_trainable: False + ucg_rate: 0.0 + input_key: concat + target: sgm.modules.encoders.modules.IdentityEncoder + + - is_trainable: False + ucg_rate: 0.0 + input_key: replace + target: sgm.modules.encoders.modules.IdentityEncoder + + - is_trainable: False + ucg_rate: 0.0 + input_key: mask + target: sgm.modules.encoders.modules.IdentityEncoder + + - is_trainable: False # crossattn + ucg_rate: 0.0 + input_keys: ["frames", "mask"] + target: sgm.modules.encoders.modules.SevaFrozenOpenCLIPImageEmbedder + params: + unsqueeze_dim: True + + first_stage_config: + target: sgm.models.autoencoder.IdentityFirstStage + + loss_fn_config: + target: sgm.modules.diffusionmodules.loss.StandardDiffusionLoss + params: + loss_weighting_config: + target: sgm.modules.diffusionmodules.loss_weighting.SevaWeighting + sigma_sampler_config: + target: sgm.modules.diffusionmodules.sigma_sampling.DiscreteSampling + params: + num_idx: 1000 + + discretization_config: + target: sgm.modules.diffusionmodules.discretizer.LegacyDDPMDiscretization + + sampler_config: + target: sgm.modules.diffusionmodules.sampling.EulerEDMSampler + params: + num_steps: 50 + + discretization_config: + target: sgm.modules.diffusionmodules.discretizer.LegacyDDPMDiscretization + + guider_config: + target: sgm.modules.diffusionmodules.guiders.VanillaCFG + params: + scale: 7.5 + +data: + target: sgm.data.dataset.DL3DVDataModuleFromConfig + params: + dataset_dir: you/path/DL3DV-ALL-480P + colmap_dir: you/path/DL3DV-ALL-ColmapCache + batch_size: 1 + num_workers: 48 + +lightning: + modelcheckpoint: + params: + every_n_train_steps: 5000 + + callbacks: + metrics_over_trainsteps_checkpoint: + params: + every_n_train_steps: 25000 + + image_logger: + target: main.ImageLogger + params: + disabled: False + enable_autocast: False + batch_frequency: 1000 + max_images: 8 + increase_log_steps: True + log_first_step: False + log_images_kwargs: + use_ema_scope: False + N: 8 + n_rows: 2 + + trainer: + devices: 0, + benchmark: True + num_sanity_val_steps: 0 + accumulate_grad_batches: 1 + max_epochs: 1000 \ No newline at end of file diff --git a/configs/example_training/toy/cifar10_cond.yaml b/configs/example_training/toy/cifar10_cond.yaml new file mode 100644 index 0000000..fca9958 --- /dev/null +++ b/configs/example_training/toy/cifar10_cond.yaml @@ -0,0 +1,98 @@ +model: + base_learning_rate: 1.0e-4 + target: sgm.models.diffusion.DiffusionEngine + params: + denoiser_config: + target: sgm.modules.diffusionmodules.denoiser.Denoiser + params: + scaling_config: + target: sgm.modules.diffusionmodules.denoiser_scaling.EDMScaling + params: + sigma_data: 1.0 + + network_config: + target: sgm.modules.diffusionmodules.openaimodel.UNetModel + params: + in_channels: 3 + out_channels: 3 + model_channels: 32 + attention_resolutions: [] + num_res_blocks: 4 + channel_mult: [1, 2, 2] + num_head_channels: 32 + num_classes: sequential + adm_in_channels: 128 + + conditioner_config: + target: sgm.modules.GeneralConditioner + params: + emb_models: + - is_trainable: True + input_key: cls + ucg_rate: 0.2 + target: sgm.modules.encoders.modules.ClassEmbedder + params: + embed_dim: 128 + n_classes: 10 + + first_stage_config: + target: sgm.models.autoencoder.IdentityFirstStage + + loss_fn_config: + target: sgm.modules.diffusionmodules.loss.StandardDiffusionLoss + params: + loss_weighting_config: + target: sgm.modules.diffusionmodules.loss_weighting.EDMWeighting + params: + sigma_data: 1.0 + sigma_sampler_config: + target: sgm.modules.diffusionmodules.sigma_sampling.EDMSampling + + sampler_config: + target: sgm.modules.diffusionmodules.sampling.EulerEDMSampler + params: + num_steps: 50 + + discretization_config: + target: sgm.modules.diffusionmodules.discretizer.EDMDiscretization + + guider_config: + target: sgm.modules.diffusionmodules.guiders.VanillaCFG + params: + scale: 3.0 + +data: + target: sgm.data.cifar10.CIFAR10Loader + params: + batch_size: 512 + num_workers: 1 + +lightning: + modelcheckpoint: + params: + every_n_train_steps: 5000 + + callbacks: + metrics_over_trainsteps_checkpoint: + params: + every_n_train_steps: 25000 + + image_logger: + target: main.ImageLogger + params: + disabled: False + batch_frequency: 1000 + max_images: 64 + increase_log_steps: True + log_first_step: False + log_images_kwargs: + use_ema_scope: False + N: 64 + n_rows: 8 + + trainer: + devices: 0, + benchmark: True + num_sanity_val_steps: 0 + accumulate_grad_batches: 1 + max_epochs: 20 \ No newline at end of file diff --git a/configs/example_training/toy/mnist.yaml b/configs/example_training/toy/mnist.yaml new file mode 100644 index 0000000..a86d05c --- /dev/null +++ b/configs/example_training/toy/mnist.yaml @@ -0,0 +1,79 @@ +model: + base_learning_rate: 1.0e-4 + target: sgm.models.diffusion.DiffusionEngine + params: + denoiser_config: + target: sgm.modules.diffusionmodules.denoiser.Denoiser + params: + scaling_config: + target: sgm.modules.diffusionmodules.denoiser_scaling.EDMScaling + params: + sigma_data: 1.0 + + network_config: + target: sgm.modules.diffusionmodules.openaimodel.UNetModel + params: + in_channels: 1 + out_channels: 1 + model_channels: 32 + attention_resolutions: [] + num_res_blocks: 4 + channel_mult: [1, 2, 2] + num_head_channels: 32 + + first_stage_config: + target: sgm.models.autoencoder.IdentityFirstStage + + loss_fn_config: + target: sgm.modules.diffusionmodules.loss.StandardDiffusionLoss + params: + loss_weighting_config: + target: sgm.modules.diffusionmodules.loss_weighting.EDMWeighting + params: + sigma_data: 1.0 + sigma_sampler_config: + target: sgm.modules.diffusionmodules.sigma_sampling.EDMSampling + + sampler_config: + target: sgm.modules.diffusionmodules.sampling.EulerEDMSampler + params: + num_steps: 50 + + discretization_config: + target: sgm.modules.diffusionmodules.discretizer.EDMDiscretization + +data: + target: sgm.data.mnist.MNISTLoader + params: + batch_size: 512 + num_workers: 1 + +lightning: + modelcheckpoint: + params: + every_n_train_steps: 5000 + + callbacks: + metrics_over_trainsteps_checkpoint: + params: + every_n_train_steps: 25000 + + image_logger: + target: main.ImageLogger + params: + disabled: False + batch_frequency: 1000 + max_images: 64 + increase_log_steps: False + log_first_step: False + log_images_kwargs: + use_ema_scope: False + N: 64 + n_rows: 8 + + trainer: + devices: 0, + benchmark: True + num_sanity_val_steps: 0 + accumulate_grad_batches: 1 + max_epochs: 10 \ No newline at end of file diff --git a/configs/example_training/toy/mnist_cond.yaml b/configs/example_training/toy/mnist_cond.yaml new file mode 100644 index 0000000..d3eafb8 --- /dev/null +++ b/configs/example_training/toy/mnist_cond.yaml @@ -0,0 +1,98 @@ +model: + base_learning_rate: 1.0e-4 + target: sgm.models.diffusion.DiffusionEngine + params: + denoiser_config: + target: sgm.modules.diffusionmodules.denoiser.Denoiser + params: + scaling_config: + target: sgm.modules.diffusionmodules.denoiser_scaling.EDMScaling + params: + sigma_data: 1.0 + + network_config: + target: sgm.modules.diffusionmodules.openaimodel.UNetModel + params: + in_channels: 1 + out_channels: 1 + model_channels: 32 + attention_resolutions: [] + num_res_blocks: 4 + channel_mult: [1, 2, 2] + num_head_channels: 32 + num_classes: sequential + adm_in_channels: 128 + + conditioner_config: + target: sgm.modules.GeneralConditioner + params: + emb_models: + - is_trainable: True + input_key: cls + ucg_rate: 0.2 + target: sgm.modules.encoders.modules.ClassEmbedder + params: + embed_dim: 128 + n_classes: 10 + + first_stage_config: + target: sgm.models.autoencoder.IdentityFirstStage + + loss_fn_config: + target: sgm.modules.diffusionmodules.loss.StandardDiffusionLoss + params: + loss_weighting_config: + target: sgm.modules.diffusionmodules.loss_weighting.EDMWeighting + params: + sigma_data: 1.0 + sigma_sampler_config: + target: sgm.modules.diffusionmodules.sigma_sampling.EDMSampling + + sampler_config: + target: sgm.modules.diffusionmodules.sampling.EulerEDMSampler + params: + num_steps: 50 + + discretization_config: + target: sgm.modules.diffusionmodules.discretizer.EDMDiscretization + + guider_config: + target: sgm.modules.diffusionmodules.guiders.VanillaCFG + params: + scale: 3.0 + +data: + target: sgm.data.mnist.MNISTLoader + params: + batch_size: 512 + num_workers: 48 + +lightning: + modelcheckpoint: + params: + every_n_train_steps: 5000 + + callbacks: + metrics_over_trainsteps_checkpoint: + params: + every_n_train_steps: 25000 + + image_logger: + target: main.ImageLogger + params: + disabled: False + batch_frequency: 1000 + max_images: 16 + increase_log_steps: True + log_first_step: False + log_images_kwargs: + use_ema_scope: False + N: 16 + n_rows: 4 + + trainer: + devices: 0, + benchmark: True + num_sanity_val_steps: 0 + accumulate_grad_batches: 1 + max_epochs: 20 \ No newline at end of file diff --git a/configs/example_training/toy/mnist_cond_discrete_eps.yaml b/configs/example_training/toy/mnist_cond_discrete_eps.yaml new file mode 100644 index 0000000..e58aae5 --- /dev/null +++ b/configs/example_training/toy/mnist_cond_discrete_eps.yaml @@ -0,0 +1,103 @@ +model: + base_learning_rate: 1.0e-4 + target: sgm.models.diffusion.DiffusionEngine + params: + denoiser_config: + target: sgm.modules.diffusionmodules.denoiser.DiscreteDenoiser + params: + num_idx: 1000 + + scaling_config: + target: sgm.modules.diffusionmodules.denoiser_scaling.EDMScaling + discretization_config: + target: sgm.modules.diffusionmodules.discretizer.LegacyDDPMDiscretization + + network_config: + target: sgm.modules.diffusionmodules.openaimodel.UNetModel + params: + in_channels: 1 + out_channels: 1 + model_channels: 32 + attention_resolutions: [] + num_res_blocks: 4 + channel_mult: [1, 2, 2] + num_head_channels: 32 + num_classes: sequential + adm_in_channels: 128 + + conditioner_config: + target: sgm.modules.GeneralConditioner + params: + emb_models: + - is_trainable: True + input_key: cls + ucg_rate: 0.2 + target: sgm.modules.encoders.modules.ClassEmbedder + params: + embed_dim: 128 + n_classes: 10 + + first_stage_config: + target: sgm.models.autoencoder.IdentityFirstStage + + loss_fn_config: + target: sgm.modules.diffusionmodules.loss.StandardDiffusionLoss + params: + loss_weighting_config: + target: sgm.modules.diffusionmodules.loss_weighting.EDMWeighting + sigma_sampler_config: + target: sgm.modules.diffusionmodules.sigma_sampling.DiscreteSampling + params: + num_idx: 1000 + + discretization_config: + target: sgm.modules.diffusionmodules.discretizer.LegacyDDPMDiscretization + + sampler_config: + target: sgm.modules.diffusionmodules.sampling.EulerEDMSampler + params: + num_steps: 50 + + discretization_config: + target: sgm.modules.diffusionmodules.discretizer.LegacyDDPMDiscretization + + guider_config: + target: sgm.modules.diffusionmodules.guiders.VanillaCFG + params: + scale: 5.0 + +data: + target: sgm.data.mnist.MNISTLoader + params: + batch_size: 512 + num_workers: 1 + +lightning: + modelcheckpoint: + params: + every_n_train_steps: 5000 + + callbacks: + metrics_over_trainsteps_checkpoint: + params: + every_n_train_steps: 25000 + + image_logger: + target: main.ImageLogger + params: + disabled: False + batch_frequency: 1000 + max_images: 16 + increase_log_steps: True + log_first_step: False + log_images_kwargs: + use_ema_scope: False + N: 16 + n_rows: 4 + + trainer: + devices: 0, + benchmark: True + num_sanity_val_steps: 0 + accumulate_grad_batches: 1 + max_epochs: 20 \ No newline at end of file diff --git a/configs/example_training/toy/mnist_cond_l1_loss.yaml b/configs/example_training/toy/mnist_cond_l1_loss.yaml new file mode 100644 index 0000000..ee2f780 --- /dev/null +++ b/configs/example_training/toy/mnist_cond_l1_loss.yaml @@ -0,0 +1,99 @@ +model: + base_learning_rate: 1.0e-4 + target: sgm.models.diffusion.DiffusionEngine + params: + denoiser_config: + target: sgm.modules.diffusionmodules.denoiser.Denoiser + params: + scaling_config: + target: sgm.modules.diffusionmodules.denoiser_scaling.EDMScaling + params: + sigma_data: 1.0 + + network_config: + target: sgm.modules.diffusionmodules.openaimodel.UNetModel + params: + in_channels: 1 + out_channels: 1 + model_channels: 32 + attention_resolutions: [] + num_res_blocks: 4 + channel_mult: [1, 2, 2] + num_head_channels: 32 + num_classes: sequential + adm_in_channels: 128 + + conditioner_config: + target: sgm.modules.GeneralConditioner + params: + emb_models: + - is_trainable: True + input_key: cls + ucg_rate: 0.2 + target: sgm.modules.encoders.modules.ClassEmbedder + params: + embed_dim: 128 + n_classes: 10 + + first_stage_config: + target: sgm.models.autoencoder.IdentityFirstStage + + loss_fn_config: + target: sgm.modules.diffusionmodules.loss.StandardDiffusionLoss + params: + loss_type: l1 + loss_weighting_config: + target: sgm.modules.diffusionmodules.loss_weighting.EDMWeighting + params: + sigma_data: 1.0 + sigma_sampler_config: + target: sgm.modules.diffusionmodules.sigma_sampling.EDMSampling + + sampler_config: + target: sgm.modules.diffusionmodules.sampling.EulerEDMSampler + params: + num_steps: 50 + + discretization_config: + target: sgm.modules.diffusionmodules.discretizer.EDMDiscretization + + guider_config: + target: sgm.modules.diffusionmodules.guiders.VanillaCFG + params: + scale: 3.0 + +data: + target: sgm.data.mnist.MNISTLoader + params: + batch_size: 512 + num_workers: 1 + +lightning: + modelcheckpoint: + params: + every_n_train_steps: 5000 + + callbacks: + metrics_over_trainsteps_checkpoint: + params: + every_n_train_steps: 25000 + + image_logger: + target: main.ImageLogger + params: + disabled: False + batch_frequency: 1000 + max_images: 64 + increase_log_steps: True + log_first_step: False + log_images_kwargs: + use_ema_scope: False + N: 64 + n_rows: 8 + + trainer: + devices: 0, + benchmark: True + num_sanity_val_steps: 0 + accumulate_grad_batches: 1 + max_epochs: 20 \ No newline at end of file diff --git a/configs/example_training/toy/mnist_cond_with_ema.yaml b/configs/example_training/toy/mnist_cond_with_ema.yaml new file mode 100644 index 0000000..c666e71 --- /dev/null +++ b/configs/example_training/toy/mnist_cond_with_ema.yaml @@ -0,0 +1,100 @@ +model: + base_learning_rate: 1.0e-4 + target: sgm.models.diffusion.DiffusionEngine + params: + use_ema: True + + denoiser_config: + target: sgm.modules.diffusionmodules.denoiser.Denoiser + params: + scaling_config: + target: sgm.modules.diffusionmodules.denoiser_scaling.EDMScaling + params: + sigma_data: 1.0 + + network_config: + target: sgm.modules.diffusionmodules.openaimodel.UNetModel + params: + in_channels: 1 + out_channels: 1 + model_channels: 32 + attention_resolutions: [] + num_res_blocks: 4 + channel_mult: [1, 2, 2] + num_head_channels: 32 + num_classes: sequential + adm_in_channels: 128 + + conditioner_config: + target: sgm.modules.GeneralConditioner + params: + emb_models: + - is_trainable: True + input_key: cls + ucg_rate: 0.2 + target: sgm.modules.encoders.modules.ClassEmbedder + params: + embed_dim: 128 + n_classes: 10 + + first_stage_config: + target: sgm.models.autoencoder.IdentityFirstStage + + loss_fn_config: + target: sgm.modules.diffusionmodules.loss.StandardDiffusionLoss + params: + loss_weighting_config: + target: sgm.modules.diffusionmodules.loss_weighting.EDMWeighting + params: + sigma_data: 1.0 + sigma_sampler_config: + target: sgm.modules.diffusionmodules.sigma_sampling.EDMSampling + + sampler_config: + target: sgm.modules.diffusionmodules.sampling.EulerEDMSampler + params: + num_steps: 50 + + discretization_config: + target: sgm.modules.diffusionmodules.discretizer.EDMDiscretization + + guider_config: + target: sgm.modules.diffusionmodules.guiders.VanillaCFG + params: + scale: 3.0 + +data: + target: sgm.data.mnist.MNISTLoader + params: + batch_size: 512 + num_workers: 1 + +lightning: + modelcheckpoint: + params: + every_n_train_steps: 5000 + + callbacks: + metrics_over_trainsteps_checkpoint: + params: + every_n_train_steps: 25000 + + image_logger: + target: main.ImageLogger + params: + disabled: False + batch_frequency: 1000 + max_images: 64 + increase_log_steps: True + log_first_step: False + log_images_kwargs: + use_ema_scope: False + N: 64 + n_rows: 8 + + trainer: + devices: 0, + benchmark: True + num_sanity_val_steps: 0 + accumulate_grad_batches: 1 + max_epochs: 20 \ No newline at end of file diff --git a/configs/example_training/txt2img-clipl-legacy-ucg-training.yaml b/configs/example_training/txt2img-clipl-legacy-ucg-training.yaml new file mode 100644 index 0000000..0f268c3 --- /dev/null +++ b/configs/example_training/txt2img-clipl-legacy-ucg-training.yaml @@ -0,0 +1,182 @@ +model: + base_learning_rate: 1.0e-4 + target: sgm.models.diffusion.DiffusionEngine + params: + scale_factor: 0.13025 + disable_first_stage_autocast: True + log_keys: + - txt + + scheduler_config: + target: sgm.lr_scheduler.LambdaLinearScheduler + params: + warm_up_steps: [10000] + cycle_lengths: [10000000000000] + f_start: [1.e-6] + f_max: [1.] + f_min: [1.] + + denoiser_config: + target: sgm.modules.diffusionmodules.denoiser.DiscreteDenoiser + params: + num_idx: 1000 + + scaling_config: + target: sgm.modules.diffusionmodules.denoiser_scaling.EpsScaling + discretization_config: + target: sgm.modules.diffusionmodules.discretizer.LegacyDDPMDiscretization + + network_config: + target: sgm.modules.diffusionmodules.openaimodel.UNetModel + params: + use_checkpoint: True + in_channels: 4 + out_channels: 4 + model_channels: 320 + attention_resolutions: [1, 2, 4] + num_res_blocks: 2 + channel_mult: [1, 2, 4, 4] + num_head_channels: 64 + num_classes: sequential + adm_in_channels: 1792 + num_heads: 1 + transformer_depth: 1 + context_dim: 768 + spatial_transformer_attn_type: softmax-xformers + + conditioner_config: + target: sgm.modules.GeneralConditioner + params: + emb_models: + - is_trainable: True + input_key: txt + ucg_rate: 0.1 + legacy_ucg_value: "" + target: sgm.modules.encoders.modules.FrozenCLIPEmbedder + params: + always_return_pooled: True + + - is_trainable: False + ucg_rate: 0.1 + input_key: original_size_as_tuple + target: sgm.modules.encoders.modules.ConcatTimestepEmbedderND + params: + outdim: 256 + + - is_trainable: False + input_key: crop_coords_top_left + ucg_rate: 0.1 + target: sgm.modules.encoders.modules.ConcatTimestepEmbedderND + params: + outdim: 256 + + first_stage_config: + target: sgm.models.autoencoder.AutoencoderKL + params: + ckpt_path: CKPT_PATH + embed_dim: 4 + monitor: val/rec_loss + ddconfig: + attn_type: vanilla-xformers + double_z: true + z_channels: 4 + resolution: 256 + in_channels: 3 + out_ch: 3 + ch: 128 + ch_mult: [ 1, 2, 4, 4 ] + num_res_blocks: 2 + attn_resolutions: [ ] + dropout: 0.0 + lossconfig: + target: torch.nn.Identity + + loss_fn_config: + target: sgm.modules.diffusionmodules.loss.StandardDiffusionLoss + params: + loss_weighting_config: + target: sgm.modules.diffusionmodules.loss_weighting.EpsWeighting + sigma_sampler_config: + target: sgm.modules.diffusionmodules.sigma_sampling.DiscreteSampling + params: + num_idx: 1000 + + discretization_config: + target: sgm.modules.diffusionmodules.discretizer.LegacyDDPMDiscretization + + sampler_config: + target: sgm.modules.diffusionmodules.sampling.EulerEDMSampler + params: + num_steps: 50 + + discretization_config: + target: sgm.modules.diffusionmodules.discretizer.LegacyDDPMDiscretization + + guider_config: + target: sgm.modules.diffusionmodules.guiders.VanillaCFG + params: + scale: 7.5 + +data: + target: sgm.data.dataset.StableDataModuleFromConfig + params: + train: + datapipeline: + urls: + # USER: adapt this path the root of your custom dataset + - DATA_PATH + pipeline_config: + shardshuffle: 10000 + sample_shuffle: 10000 # USER: you might wanna adapt depending on your available RAM + + decoders: + - pil + + postprocessors: + - target: sdata.mappers.TorchVisionImageTransforms + params: + key: jpg # USER: you might wanna adapt this for your custom dataset + transforms: + - target: torchvision.transforms.Resize + params: + size: 256 + interpolation: 3 + - target: torchvision.transforms.ToTensor + - target: sdata.mappers.Rescaler + - target: sdata.mappers.AddOriginalImageSizeAsTupleAndCropToSquare + # USER: you might wanna use non-default parameters due to your custom dataset + + loader: + batch_size: 64 + num_workers: 6 + +lightning: + modelcheckpoint: + params: + every_n_train_steps: 5000 + + callbacks: + metrics_over_trainsteps_checkpoint: + params: + every_n_train_steps: 25000 + + image_logger: + target: main.ImageLogger + params: + disabled: False + enable_autocast: False + batch_frequency: 1000 + max_images: 8 + increase_log_steps: True + log_first_step: False + log_images_kwargs: + use_ema_scope: False + N: 8 + n_rows: 2 + + trainer: + devices: 0, + benchmark: True + num_sanity_val_steps: 0 + accumulate_grad_batches: 1 + max_epochs: 1000 \ No newline at end of file diff --git a/configs/example_training/txt2img-clipl.yaml b/configs/example_training/txt2img-clipl.yaml new file mode 100644 index 0000000..3de0e8f --- /dev/null +++ b/configs/example_training/txt2img-clipl.yaml @@ -0,0 +1,184 @@ +model: + base_learning_rate: 1.0e-4 + target: sgm.models.diffusion.DiffusionEngine + params: + scale_factor: 0.13025 + disable_first_stage_autocast: True + log_keys: + - txt + + scheduler_config: + target: sgm.lr_scheduler.LambdaLinearScheduler + params: + warm_up_steps: [10000] + cycle_lengths: [10000000000000] + f_start: [1.e-6] + f_max: [1.] + f_min: [1.] + + denoiser_config: + target: sgm.modules.diffusionmodules.denoiser.DiscreteDenoiser + params: + num_idx: 1000 + + scaling_config: + target: sgm.modules.diffusionmodules.denoiser_scaling.EpsScaling + discretization_config: + target: sgm.modules.diffusionmodules.discretizer.LegacyDDPMDiscretization + + network_config: + target: sgm.modules.diffusionmodules.openaimodel.UNetModel + params: + use_checkpoint: True + in_channels: 4 + out_channels: 4 + model_channels: 320 + attention_resolutions: [1, 2, 4] + num_res_blocks: 2 + channel_mult: [1, 2, 4, 4] + num_head_channels: 64 + num_classes: sequential + adm_in_channels: 1792 + num_heads: 1 + transformer_depth: 1 + context_dim: 768 + spatial_transformer_attn_type: softmax-xformers + + conditioner_config: + target: sgm.modules.GeneralConditioner + params: + emb_models: + - is_trainable: True + input_key: txt + ucg_rate: 0.1 + legacy_ucg_value: "" + target: sgm.modules.encoders.modules.FrozenCLIPEmbedder + params: + always_return_pooled: True + + - is_trainable: False + ucg_rate: 0.1 + input_key: original_size_as_tuple + target: sgm.modules.encoders.modules.ConcatTimestepEmbedderND + params: + outdim: 256 + + - is_trainable: False + input_key: crop_coords_top_left + ucg_rate: 0.1 + target: sgm.modules.encoders.modules.ConcatTimestepEmbedderND + params: + outdim: 256 + + first_stage_config: + target: sgm.models.autoencoder.AutoencoderKL + params: + ckpt_path: + embed_dim: 4 + monitor: val/rec_loss + ddconfig: + attn_type: vanilla-xformers + double_z: true + z_channels: 4 + resolution: 256 + in_channels: 3 + out_ch: 3 + ch: 128 + ch_mult: [1, 2, 4, 4] + num_res_blocks: 2 + attn_resolutions: [] + dropout: 0.0 + lossconfig: + target: torch.nn.Identity + + loss_fn_config: + target: sgm.modules.diffusionmodules.loss.StandardDiffusionLoss + params: + loss_weighting_config: + target: sgm.modules.diffusionmodules.loss_weighting.EpsWeighting + sigma_sampler_config: + target: sgm.modules.diffusionmodules.sigma_sampling.DiscreteSampling + params: + num_idx: 1000 + + discretization_config: + target: sgm.modules.diffusionmodules.discretizer.LegacyDDPMDiscretization + + sampler_config: + target: sgm.modules.diffusionmodules.sampling.EulerEDMSampler + params: + num_steps: 50 + + discretization_config: + target: sgm.modules.diffusionmodules.discretizer.LegacyDDPMDiscretization + + guider_config: + target: sgm.modules.diffusionmodules.guiders.VanillaCFG + params: + scale: 7.5 + +data: + target: sgm.data.dataset.StableDataModuleFromConfig + params: + train: + datapipeline: + urls: + # USER: adapt this path the root of your custom dataset + - DATA_PATH + pipeline_config: + shardshuffle: 10000 + sample_shuffle: 10000 + + + decoders: + - pil + + postprocessors: + - target: sdata.mappers.TorchVisionImageTransforms + params: + key: jpg # USER: you might wanna adapt this for your custom dataset + transforms: + - target: torchvision.transforms.Resize + params: + size: 256 + interpolation: 3 + - target: torchvision.transforms.ToTensor + - target: sdata.mappers.Rescaler + # USER: you might wanna use non-default parameters due to your custom dataset + - target: sdata.mappers.AddOriginalImageSizeAsTupleAndCropToSquare + # USER: you might wanna use non-default parameters due to your custom dataset + + loader: + batch_size: 64 + num_workers: 6 + +lightning: + modelcheckpoint: + params: + every_n_train_steps: 5000 + + callbacks: + metrics_over_trainsteps_checkpoint: + params: + every_n_train_steps: 25000 + + image_logger: + target: main.ImageLogger + params: + disabled: False + enable_autocast: False + batch_frequency: 1000 + max_images: 8 + increase_log_steps: True + log_first_step: False + log_images_kwargs: + use_ema_scope: False + N: 8 + n_rows: 2 + + trainer: + devices: 0, + benchmark: True + num_sanity_val_steps: 0 + accumulate_grad_batches: 1 + max_epochs: 1000 \ No newline at end of file diff --git a/configs/inference/sd_2_1.yaml b/configs/inference/sd_2_1.yaml new file mode 100644 index 0000000..6531c6c --- /dev/null +++ b/configs/inference/sd_2_1.yaml @@ -0,0 +1,60 @@ +model: + target: sgm.models.diffusion.DiffusionEngine + params: + scale_factor: 0.18215 + disable_first_stage_autocast: True + + denoiser_config: + target: sgm.modules.diffusionmodules.denoiser.DiscreteDenoiser + params: + num_idx: 1000 + + scaling_config: + target: sgm.modules.diffusionmodules.denoiser_scaling.EpsScaling + discretization_config: + target: sgm.modules.diffusionmodules.discretizer.LegacyDDPMDiscretization + + network_config: + target: sgm.modules.diffusionmodules.openaimodel.UNetModel + params: + use_checkpoint: True + in_channels: 4 + out_channels: 4 + model_channels: 320 + attention_resolutions: [4, 2, 1] + num_res_blocks: 2 + channel_mult: [1, 2, 4, 4] + num_head_channels: 64 + use_linear_in_transformer: True + transformer_depth: 1 + context_dim: 1024 + + conditioner_config: + target: sgm.modules.GeneralConditioner + params: + emb_models: + - is_trainable: False + input_key: txt + target: sgm.modules.encoders.modules.FrozenOpenCLIPEmbedder + params: + freeze: true + layer: penultimate + + first_stage_config: + target: sgm.models.autoencoder.AutoencoderKL + params: + embed_dim: 4 + monitor: val/rec_loss + ddconfig: + double_z: true + z_channels: 4 + resolution: 256 + in_channels: 3 + out_ch: 3 + ch: 128 + ch_mult: [1, 2, 4, 4] + num_res_blocks: 2 + attn_resolutions: [] + dropout: 0.0 + lossconfig: + target: torch.nn.Identity \ No newline at end of file diff --git a/configs/inference/sd_2_1_768.yaml b/configs/inference/sd_2_1_768.yaml new file mode 100644 index 0000000..e2f9910 --- /dev/null +++ b/configs/inference/sd_2_1_768.yaml @@ -0,0 +1,60 @@ +model: + target: sgm.models.diffusion.DiffusionEngine + params: + scale_factor: 0.18215 + disable_first_stage_autocast: True + + denoiser_config: + target: sgm.modules.diffusionmodules.denoiser.DiscreteDenoiser + params: + num_idx: 1000 + + scaling_config: + target: sgm.modules.diffusionmodules.denoiser_scaling.VScaling + discretization_config: + target: sgm.modules.diffusionmodules.discretizer.LegacyDDPMDiscretization + + network_config: + target: sgm.modules.diffusionmodules.openaimodel.UNetModel + params: + use_checkpoint: True + in_channels: 4 + out_channels: 4 + model_channels: 320 + attention_resolutions: [4, 2, 1] + num_res_blocks: 2 + channel_mult: [1, 2, 4, 4] + num_head_channels: 64 + use_linear_in_transformer: True + transformer_depth: 1 + context_dim: 1024 + + conditioner_config: + target: sgm.modules.GeneralConditioner + params: + emb_models: + - is_trainable: False + input_key: txt + target: sgm.modules.encoders.modules.FrozenOpenCLIPEmbedder + params: + freeze: true + layer: penultimate + + first_stage_config: + target: sgm.models.autoencoder.AutoencoderKL + params: + embed_dim: 4 + monitor: val/rec_loss + ddconfig: + double_z: true + z_channels: 4 + resolution: 256 + in_channels: 3 + out_ch: 3 + ch: 128 + ch_mult: [1, 2, 4, 4] + num_res_blocks: 2 + attn_resolutions: [] + dropout: 0.0 + lossconfig: + target: torch.nn.Identity \ No newline at end of file diff --git a/configs/inference/sd_xl_base.yaml b/configs/inference/sd_xl_base.yaml new file mode 100644 index 0000000..6047379 --- /dev/null +++ b/configs/inference/sd_xl_base.yaml @@ -0,0 +1,93 @@ +model: + target: sgm.models.diffusion.DiffusionEngine + params: + scale_factor: 0.13025 + disable_first_stage_autocast: True + + denoiser_config: + target: sgm.modules.diffusionmodules.denoiser.DiscreteDenoiser + params: + num_idx: 1000 + + scaling_config: + target: sgm.modules.diffusionmodules.denoiser_scaling.EpsScaling + discretization_config: + target: sgm.modules.diffusionmodules.discretizer.LegacyDDPMDiscretization + + network_config: + target: sgm.modules.diffusionmodules.openaimodel.UNetModel + params: + adm_in_channels: 2816 + num_classes: sequential + use_checkpoint: True + in_channels: 4 + out_channels: 4 + model_channels: 320 + attention_resolutions: [4, 2] + num_res_blocks: 2 + channel_mult: [1, 2, 4] + num_head_channels: 64 + use_linear_in_transformer: True + transformer_depth: [1, 2, 10] + context_dim: 2048 + spatial_transformer_attn_type: softmax-xformers + + conditioner_config: + target: sgm.modules.GeneralConditioner + params: + emb_models: + - is_trainable: False + input_key: txt + target: sgm.modules.encoders.modules.FrozenCLIPEmbedder + params: + layer: hidden + layer_idx: 11 + + - is_trainable: False + input_key: txt + target: sgm.modules.encoders.modules.FrozenOpenCLIPEmbedder2 + params: + arch: ViT-bigG-14 + version: laion2b_s39b_b160k + freeze: True + layer: penultimate + always_return_pooled: True + legacy: False + + - is_trainable: False + input_key: original_size_as_tuple + target: sgm.modules.encoders.modules.ConcatTimestepEmbedderND + params: + outdim: 256 + + - is_trainable: False + input_key: crop_coords_top_left + target: sgm.modules.encoders.modules.ConcatTimestepEmbedderND + params: + outdim: 256 + + - is_trainable: False + input_key: target_size_as_tuple + target: sgm.modules.encoders.modules.ConcatTimestepEmbedderND + params: + outdim: 256 + + first_stage_config: + target: sgm.models.autoencoder.AutoencoderKL + params: + embed_dim: 4 + monitor: val/rec_loss + ddconfig: + attn_type: vanilla-xformers + double_z: true + z_channels: 4 + resolution: 256 + in_channels: 3 + out_ch: 3 + ch: 128 + ch_mult: [1, 2, 4, 4] + num_res_blocks: 2 + attn_resolutions: [] + dropout: 0.0 + lossconfig: + target: torch.nn.Identity diff --git a/configs/inference/sd_xl_refiner.yaml b/configs/inference/sd_xl_refiner.yaml new file mode 100644 index 0000000..2d5ab44 --- /dev/null +++ b/configs/inference/sd_xl_refiner.yaml @@ -0,0 +1,86 @@ +model: + target: sgm.models.diffusion.DiffusionEngine + params: + scale_factor: 0.13025 + disable_first_stage_autocast: True + + denoiser_config: + target: sgm.modules.diffusionmodules.denoiser.DiscreteDenoiser + params: + num_idx: 1000 + + scaling_config: + target: sgm.modules.diffusionmodules.denoiser_scaling.EpsScaling + discretization_config: + target: sgm.modules.diffusionmodules.discretizer.LegacyDDPMDiscretization + + network_config: + target: sgm.modules.diffusionmodules.openaimodel.UNetModel + params: + adm_in_channels: 2560 + num_classes: sequential + use_checkpoint: True + in_channels: 4 + out_channels: 4 + model_channels: 384 + attention_resolutions: [4, 2] + num_res_blocks: 2 + channel_mult: [1, 2, 4, 4] + num_head_channels: 64 + use_linear_in_transformer: True + transformer_depth: 4 + context_dim: [1280, 1280, 1280, 1280] + spatial_transformer_attn_type: softmax-xformers + + conditioner_config: + target: sgm.modules.GeneralConditioner + params: + emb_models: + - is_trainable: False + input_key: txt + target: sgm.modules.encoders.modules.FrozenOpenCLIPEmbedder2 + params: + arch: ViT-bigG-14 + version: laion2b_s39b_b160k + legacy: False + freeze: True + layer: penultimate + always_return_pooled: True + + - is_trainable: False + input_key: original_size_as_tuple + target: sgm.modules.encoders.modules.ConcatTimestepEmbedderND + params: + outdim: 256 + + - is_trainable: False + input_key: crop_coords_top_left + target: sgm.modules.encoders.modules.ConcatTimestepEmbedderND + params: + outdim: 256 + + - is_trainable: False + input_key: aesthetic_score + target: sgm.modules.encoders.modules.ConcatTimestepEmbedderND + params: + outdim: 256 + + first_stage_config: + target: sgm.models.autoencoder.AutoencoderKL + params: + embed_dim: 4 + monitor: val/rec_loss + ddconfig: + attn_type: vanilla-xformers + double_z: true + z_channels: 4 + resolution: 256 + in_channels: 3 + out_ch: 3 + ch: 128 + ch_mult: [1, 2, 4, 4] + num_res_blocks: 2 + attn_resolutions: [] + dropout: 0.0 + lossconfig: + target: torch.nn.Identity diff --git a/configs/inference/sv3d_p.yaml b/configs/inference/sv3d_p.yaml new file mode 100644 index 0000000..d3781fe --- /dev/null +++ b/configs/inference/sv3d_p.yaml @@ -0,0 +1,118 @@ +model: + target: sgm.models.diffusion.DiffusionEngine + params: + scale_factor: 0.18215 + disable_first_stage_autocast: True + + denoiser_config: + target: sgm.modules.diffusionmodules.denoiser.Denoiser + params: + scaling_config: + target: sgm.modules.diffusionmodules.denoiser_scaling.VScalingWithEDMcNoise + + network_config: + target: sgm.modules.diffusionmodules.video_model.VideoUNet + params: + adm_in_channels: 1280 + num_classes: sequential + use_checkpoint: True + in_channels: 8 + out_channels: 4 + model_channels: 320 + attention_resolutions: [4, 2, 1] + num_res_blocks: 2 + channel_mult: [1, 2, 4, 4] + num_head_channels: 64 + use_linear_in_transformer: True + transformer_depth: 1 + context_dim: 1024 + spatial_transformer_attn_type: softmax-xformers + extra_ff_mix_layer: True + use_spatial_context: True + merge_strategy: learned_with_images + video_kernel_size: [3, 1, 1] + + conditioner_config: + target: sgm.modules.GeneralConditioner + params: + emb_models: + - input_key: cond_frames_without_noise + is_trainable: False + target: sgm.modules.encoders.modules.FrozenOpenCLIPImagePredictionEmbedder + params: + n_cond_frames: 1 + n_copies: 1 + open_clip_embedding_config: + target: sgm.modules.encoders.modules.FrozenOpenCLIPImageEmbedder + params: + freeze: True + + - input_key: cond_frames + is_trainable: False + target: sgm.modules.encoders.modules.VideoPredictionEmbedderWithEncoder + params: + disable_encoder_autocast: True + n_cond_frames: 1 + n_copies: 1 + is_ae: True + encoder_config: + target: sgm.models.autoencoder.AutoencoderKLModeOnly + params: + embed_dim: 4 + monitor: val/rec_loss + ddconfig: + attn_type: vanilla-xformers + double_z: True + z_channels: 4 + resolution: 256 + in_channels: 3 + out_ch: 3 + ch: 128 + ch_mult: [1, 2, 4, 4] + num_res_blocks: 2 + attn_resolutions: [] + dropout: 0.0 + lossconfig: + target: torch.nn.Identity + + - input_key: cond_aug + is_trainable: False + target: sgm.modules.encoders.modules.ConcatTimestepEmbedderND + params: + outdim: 256 + + - input_key: polars_rad + is_trainable: False + target: sgm.modules.encoders.modules.ConcatTimestepEmbedderND + params: + outdim: 512 + + - input_key: azimuths_rad + is_trainable: False + target: sgm.modules.encoders.modules.ConcatTimestepEmbedderND + params: + outdim: 512 + + first_stage_config: + target: sgm.models.autoencoder.AutoencodingEngine + params: + loss_config: + target: torch.nn.Identity + regularizer_config: + target: sgm.modules.autoencoding.regularizers.DiagonalGaussianRegularizer + encoder_config: + target: torch.nn.Identity + decoder_config: + target: sgm.modules.diffusionmodules.model.Decoder + params: + attn_type: vanilla-xformers + double_z: True + z_channels: 4 + resolution: 256 + in_channels: 3 + out_ch: 3 + ch: 128 + ch_mult: [ 1, 2, 4, 4 ] + num_res_blocks: 2 + attn_resolutions: [ ] + dropout: 0.0 \ No newline at end of file diff --git a/configs/inference/sv3d_u.yaml b/configs/inference/sv3d_u.yaml new file mode 100644 index 0000000..5c48a5f --- /dev/null +++ b/configs/inference/sv3d_u.yaml @@ -0,0 +1,106 @@ +model: + target: sgm.models.diffusion.DiffusionEngine + params: + scale_factor: 0.18215 + disable_first_stage_autocast: True + + denoiser_config: + target: sgm.modules.diffusionmodules.denoiser.Denoiser + params: + scaling_config: + target: sgm.modules.diffusionmodules.denoiser_scaling.VScalingWithEDMcNoise + + network_config: + target: sgm.modules.diffusionmodules.video_model.VideoUNet + params: + adm_in_channels: 256 + num_classes: sequential + use_checkpoint: True + in_channels: 8 + out_channels: 4 + model_channels: 320 + attention_resolutions: [4, 2, 1] + num_res_blocks: 2 + channel_mult: [1, 2, 4, 4] + num_head_channels: 64 + use_linear_in_transformer: True + transformer_depth: 1 + context_dim: 1024 + spatial_transformer_attn_type: softmax-xformers + extra_ff_mix_layer: True + use_spatial_context: True + merge_strategy: learned_with_images + video_kernel_size: [3, 1, 1] + + conditioner_config: + target: sgm.modules.GeneralConditioner + params: + emb_models: + - input_key: cond_frames_without_noise + is_trainable: False + target: sgm.modules.encoders.modules.FrozenOpenCLIPImagePredictionEmbedder + params: + n_cond_frames: 1 + n_copies: 1 + open_clip_embedding_config: + target: sgm.modules.encoders.modules.FrozenOpenCLIPImageEmbedder + params: + freeze: True + + - input_key: cond_frames + is_trainable: False + target: sgm.modules.encoders.modules.VideoPredictionEmbedderWithEncoder + params: + disable_encoder_autocast: True + n_cond_frames: 1 + n_copies: 1 + is_ae: True + encoder_config: + target: sgm.models.autoencoder.AutoencoderKLModeOnly + params: + embed_dim: 4 + monitor: val/rec_loss + ddconfig: + attn_type: vanilla-xformers + double_z: True + z_channels: 4 + resolution: 256 + in_channels: 3 + out_ch: 3 + ch: 128 + ch_mult: [1, 2, 4, 4] + num_res_blocks: 2 + attn_resolutions: [] + dropout: 0.0 + lossconfig: + target: torch.nn.Identity + + - input_key: cond_aug + is_trainable: False + target: sgm.modules.encoders.modules.ConcatTimestepEmbedderND + params: + outdim: 256 + + first_stage_config: + target: sgm.models.autoencoder.AutoencodingEngine + params: + loss_config: + target: torch.nn.Identity + regularizer_config: + target: sgm.modules.autoencoding.regularizers.DiagonalGaussianRegularizer + encoder_config: + target: torch.nn.Identity + decoder_config: + target: sgm.modules.diffusionmodules.model.Decoder + params: + attn_type: vanilla-xformers + double_z: True + z_channels: 4 + resolution: 256 + in_channels: 3 + out_ch: 3 + ch: 128 + ch_mult: [ 1, 2, 4, 4 ] + num_res_blocks: 2 + attn_resolutions: [ ] + dropout: 0.0 \ No newline at end of file diff --git a/configs/inference/svd.yaml b/configs/inference/svd.yaml new file mode 100644 index 0000000..2a0819e --- /dev/null +++ b/configs/inference/svd.yaml @@ -0,0 +1,131 @@ +model: + target: sgm.models.diffusion.DiffusionEngine + params: + scale_factor: 0.18215 + disable_first_stage_autocast: True + + denoiser_config: + target: sgm.modules.diffusionmodules.denoiser.Denoiser + params: + scaling_config: + target: sgm.modules.diffusionmodules.denoiser_scaling.VScalingWithEDMcNoise + + network_config: + target: sgm.modules.diffusionmodules.video_model.VideoUNet + params: + adm_in_channels: 768 + num_classes: sequential + use_checkpoint: True + in_channels: 8 + out_channels: 4 + model_channels: 320 + attention_resolutions: [4, 2, 1] + num_res_blocks: 2 + channel_mult: [1, 2, 4, 4] + num_head_channels: 64 + use_linear_in_transformer: True + transformer_depth: 1 + context_dim: 1024 + spatial_transformer_attn_type: softmax-xformers + extra_ff_mix_layer: True + use_spatial_context: True + merge_strategy: learned_with_images + video_kernel_size: [3, 1, 1] + + conditioner_config: + target: sgm.modules.GeneralConditioner + params: + emb_models: + - is_trainable: False + input_key: cond_frames_without_noise + target: sgm.modules.encoders.modules.FrozenOpenCLIPImagePredictionEmbedder + params: + n_cond_frames: 1 + n_copies: 1 + open_clip_embedding_config: + target: sgm.modules.encoders.modules.FrozenOpenCLIPImageEmbedder + params: + freeze: True + + - input_key: fps_id + is_trainable: False + target: sgm.modules.encoders.modules.ConcatTimestepEmbedderND + params: + outdim: 256 + + - input_key: motion_bucket_id + is_trainable: False + target: sgm.modules.encoders.modules.ConcatTimestepEmbedderND + params: + outdim: 256 + + - input_key: cond_frames + is_trainable: False + target: sgm.modules.encoders.modules.VideoPredictionEmbedderWithEncoder + params: + disable_encoder_autocast: True + n_cond_frames: 1 + n_copies: 1 + is_ae: True + encoder_config: + target: sgm.models.autoencoder.AutoencoderKLModeOnly + params: + embed_dim: 4 + monitor: val/rec_loss + ddconfig: + attn_type: vanilla-xformers + double_z: True + z_channels: 4 + resolution: 256 + in_channels: 3 + out_ch: 3 + ch: 128 + ch_mult: [1, 2, 4, 4] + num_res_blocks: 2 + attn_resolutions: [] + dropout: 0.0 + lossconfig: + target: torch.nn.Identity + + - input_key: cond_aug + is_trainable: False + target: sgm.modules.encoders.modules.ConcatTimestepEmbedderND + params: + outdim: 256 + + first_stage_config: + target: sgm.models.autoencoder.AutoencodingEngine + params: + loss_config: + target: torch.nn.Identity + regularizer_config: + target: sgm.modules.autoencoding.regularizers.DiagonalGaussianRegularizer + encoder_config: + target: sgm.modules.diffusionmodules.model.Encoder + params: + attn_type: vanilla + double_z: True + z_channels: 4 + resolution: 256 + in_channels: 3 + out_ch: 3 + ch: 128 + ch_mult: [1, 2, 4, 4] + num_res_blocks: 2 + attn_resolutions: [] + dropout: 0.0 + decoder_config: + target: sgm.modules.autoencoding.temporal_ae.VideoDecoder + params: + attn_type: vanilla + double_z: True + z_channels: 4 + resolution: 256 + in_channels: 3 + out_ch: 3 + ch: 128 + ch_mult: [1, 2, 4, 4] + num_res_blocks: 2 + attn_resolutions: [] + dropout: 0.0 + video_kernel_size: [3, 1, 1] \ No newline at end of file diff --git a/configs/inference/svd_image_decoder.yaml b/configs/inference/svd_image_decoder.yaml new file mode 100644 index 0000000..bb09177 --- /dev/null +++ b/configs/inference/svd_image_decoder.yaml @@ -0,0 +1,114 @@ +model: + target: sgm.models.diffusion.DiffusionEngine + params: + scale_factor: 0.18215 + disable_first_stage_autocast: True + + denoiser_config: + target: sgm.modules.diffusionmodules.denoiser.Denoiser + params: + scaling_config: + target: sgm.modules.diffusionmodules.denoiser_scaling.VScalingWithEDMcNoise + + network_config: + target: sgm.modules.diffusionmodules.video_model.VideoUNet + params: + adm_in_channels: 768 + num_classes: sequential + use_checkpoint: True + in_channels: 8 + out_channels: 4 + model_channels: 320 + attention_resolutions: [4, 2, 1] + num_res_blocks: 2 + channel_mult: [1, 2, 4, 4] + num_head_channels: 64 + use_linear_in_transformer: True + transformer_depth: 1 + context_dim: 1024 + spatial_transformer_attn_type: softmax-xformers + extra_ff_mix_layer: True + use_spatial_context: True + merge_strategy: learned_with_images + video_kernel_size: [3, 1, 1] + + conditioner_config: + target: sgm.modules.GeneralConditioner + params: + emb_models: + - is_trainable: False + input_key: cond_frames_without_noise + target: sgm.modules.encoders.modules.FrozenOpenCLIPImagePredictionEmbedder + params: + n_cond_frames: 1 + n_copies: 1 + open_clip_embedding_config: + target: sgm.modules.encoders.modules.FrozenOpenCLIPImageEmbedder + params: + freeze: True + + - input_key: fps_id + is_trainable: False + target: sgm.modules.encoders.modules.ConcatTimestepEmbedderND + params: + outdim: 256 + + - input_key: motion_bucket_id + is_trainable: False + target: sgm.modules.encoders.modules.ConcatTimestepEmbedderND + params: + outdim: 256 + + - input_key: cond_frames + is_trainable: False + target: sgm.modules.encoders.modules.VideoPredictionEmbedderWithEncoder + params: + disable_encoder_autocast: True + n_cond_frames: 1 + n_copies: 1 + is_ae: True + encoder_config: + target: sgm.models.autoencoder.AutoencoderKLModeOnly + params: + embed_dim: 4 + monitor: val/rec_loss + ddconfig: + attn_type: vanilla-xformers + double_z: True + z_channels: 4 + resolution: 256 + in_channels: 3 + out_ch: 3 + ch: 128 + ch_mult: [1, 2, 4, 4] + num_res_blocks: 2 + attn_resolutions: [] + dropout: 0.0 + lossconfig: + target: torch.nn.Identity + + - input_key: cond_aug + is_trainable: False + target: sgm.modules.encoders.modules.ConcatTimestepEmbedderND + params: + outdim: 256 + + first_stage_config: + target: sgm.models.autoencoder.AutoencoderKL + params: + embed_dim: 4 + monitor: val/rec_loss + ddconfig: + attn_type: vanilla-xformers + double_z: True + z_channels: 4 + resolution: 256 + in_channels: 3 + out_ch: 3 + ch: 128 + ch_mult: [1, 2, 4, 4] + num_res_blocks: 2 + attn_resolutions: [] + dropout: 0.0 + lossconfig: + target: torch.nn.Identity \ No newline at end of file diff --git a/main.py b/main.py new file mode 100644 index 0000000..96f2757 --- /dev/null +++ b/main.py @@ -0,0 +1,946 @@ +import argparse +import datetime +import glob +import inspect +import os +import sys +from inspect import Parameter +from typing import Union + +import numpy as np +import pytorch_lightning as pl +import torch +import torchvision +import wandb +from matplotlib import pyplot as plt +from natsort import natsorted +from omegaconf import OmegaConf +from packaging import version +from PIL import Image +from pytorch_lightning import seed_everything +from pytorch_lightning.callbacks import Callback +from pytorch_lightning.loggers import WandbLogger +from pytorch_lightning.trainer import Trainer +from pytorch_lightning.utilities import rank_zero_only + +from sgm.util import exists, instantiate_from_config, isheatmap + +MULTINODE_HACKS = True + + +def default_trainer_args(): + argspec = dict(inspect.signature(Trainer.__init__).parameters) + argspec.pop("self") + default_args = { + param: argspec[param].default + for param in argspec + if argspec[param] != Parameter.empty + } + return default_args + + +def get_parser(**parser_kwargs): + def str2bool(v): + if isinstance(v, bool): + return v + if v.lower() in ("yes", "true", "t", "y", "1"): + return True + elif v.lower() in ("no", "false", "f", "n", "0"): + return False + else: + raise argparse.ArgumentTypeError("Boolean value expected.") + + parser = argparse.ArgumentParser(**parser_kwargs) + parser.add_argument( + "-n", + "--name", + type=str, + const=True, + default="", + nargs="?", + help="postfix for logdir", + ) + parser.add_argument( + "--no_date", + type=str2bool, + nargs="?", + const=True, + default=False, + help="if True, skip date generation for logdir and only use naming via opt.base or opt.name (+ opt.postfix, optionally)", + ) + parser.add_argument( + "-r", + "--resume", + type=str, + const=True, + default="", + nargs="?", + help="resume from logdir or checkpoint in logdir", + ) + parser.add_argument( + "-b", + "--base", + nargs="*", + metavar="base_config.yaml", + help="paths to base configs. Loaded from left-to-right. " + "Parameters can be overwritten or added with command-line options of the form `--key value`.", + default=list(), + ) + parser.add_argument( + "-t", + "--train", + type=str2bool, + const=True, + default=True, + nargs="?", + help="train", + ) + parser.add_argument( + "--no-test", + type=str2bool, + const=True, + default=False, + nargs="?", + help="disable test", + ) + parser.add_argument( + "-p", "--project", help="name of new or path to existing project" + ) + parser.add_argument( + "-d", + "--debug", + type=str2bool, + nargs="?", + const=True, + default=False, + help="enable post-mortem debugging", + ) + parser.add_argument( + "-s", + "--seed", + type=int, + default=23, + help="seed for seed_everything", + ) + parser.add_argument( + "-f", + "--postfix", + type=str, + default="", + help="post-postfix for default name", + ) + parser.add_argument( + "--projectname", + type=str, + default="stablediffusion", + ) + parser.add_argument( + "-l", + "--logdir", + type=str, + default="logs", + help="directory for logging dat shit", + ) + parser.add_argument( + "--scale_lr", + type=str2bool, + nargs="?", + const=True, + default=False, + help="scale base-lr by ngpu * batch_size * n_accumulate", + ) + parser.add_argument( + "--legacy_naming", + type=str2bool, + nargs="?", + const=True, + default=False, + help="name run based on config file name if true, else by whole path", + ) + parser.add_argument( + "--enable_tf32", + type=str2bool, + nargs="?", + const=True, + default=False, + help="enables the TensorFloat32 format both for matmuls and cuDNN for pytorch 1.12", + ) + parser.add_argument( + "--startup", + type=str, + default=None, + help="Startuptime from distributed script", + ) + parser.add_argument( + "--wandb", + type=str2bool, + nargs="?", + const=True, + default=False, # TODO: later default to True + help="log to wandb", + ) + parser.add_argument( + "--no_base_name", + type=str2bool, + nargs="?", + const=True, + default=False, # TODO: later default to True + help="log to wandb", + ) + if version.parse(torch.__version__) >= version.parse("2.0.0"): + parser.add_argument( + "--resume_from_checkpoint", + type=str, + default=None, + help="single checkpoint file to resume from", + ) + default_args = default_trainer_args() + for key in default_args: + parser.add_argument("--" + key, default=default_args[key]) + return parser + + +def get_checkpoint_name(logdir): + ckpt = os.path.join(logdir, "checkpoints", "last**.ckpt") + ckpt = natsorted(glob.glob(ckpt)) + print('available "last" checkpoints:') + print(ckpt) + if len(ckpt) > 1: + print("got most recent checkpoint") + ckpt = sorted(ckpt, key=lambda x: os.path.getmtime(x))[-1] + print(f"Most recent ckpt is {ckpt}") + with open(os.path.join(logdir, "most_recent_ckpt.txt"), "w") as f: + f.write(ckpt + "\n") + try: + version = int(ckpt.split("/")[-1].split("-v")[-1].split(".")[0]) + except Exception as e: + print("version confusion but not bad") + print(e) + version = 1 + # version = last_version + 1 + else: + # in this case, we only have one "last.ckpt" + ckpt = ckpt[0] + version = 1 + melk_ckpt_name = f"last-v{version}.ckpt" + print(f"Current melk ckpt name: {melk_ckpt_name}") + return ckpt, melk_ckpt_name + + +class SetupCallback(Callback): + def __init__( + self, + resume, + now, + logdir, + ckptdir, + cfgdir, + config, + lightning_config, + debug, + ckpt_name=None, + ): + super().__init__() + self.resume = resume + self.now = now + self.logdir = logdir + self.ckptdir = ckptdir + self.cfgdir = cfgdir + self.config = config + self.lightning_config = lightning_config + self.debug = debug + self.ckpt_name = ckpt_name + + def on_exception(self, trainer: pl.Trainer, pl_module, exception): + if not self.debug and trainer.global_rank == 0: + print("Summoning checkpoint.") + if self.ckpt_name is None: + ckpt_path = os.path.join(self.ckptdir, "last.ckpt") + else: + ckpt_path = os.path.join(self.ckptdir, self.ckpt_name) + trainer.save_checkpoint(ckpt_path) + + def on_fit_start(self, trainer, pl_module): + if trainer.global_rank == 0: + # Create logdirs and save configs + os.makedirs(self.logdir, exist_ok=True) + os.makedirs(self.ckptdir, exist_ok=True) + os.makedirs(self.cfgdir, exist_ok=True) + + if "callbacks" in self.lightning_config: + if ( + "metrics_over_trainsteps_checkpoint" + in self.lightning_config["callbacks"] + ): + os.makedirs( + os.path.join(self.ckptdir, "trainstep_checkpoints"), + exist_ok=True, + ) + print("Project config") + print(OmegaConf.to_yaml(self.config)) + if MULTINODE_HACKS: + import time + + time.sleep(5) + OmegaConf.save( + self.config, + os.path.join(self.cfgdir, "{}-project.yaml".format(self.now)), + ) + + print("Lightning config") + print(OmegaConf.to_yaml(self.lightning_config)) + OmegaConf.save( + OmegaConf.create({"lightning": self.lightning_config}), + os.path.join(self.cfgdir, "{}-lightning.yaml".format(self.now)), + ) + + else: + # ModelCheckpoint callback created log directory --- remove it + if not MULTINODE_HACKS and not self.resume and os.path.exists(self.logdir): + dst, name = os.path.split(self.logdir) + dst = os.path.join(dst, "child_runs", name) + os.makedirs(os.path.split(dst)[0], exist_ok=True) + try: + os.rename(self.logdir, dst) + except FileNotFoundError: + pass + + +class ImageLogger(Callback): + def __init__( + self, + batch_frequency, + max_images, + clamp=True, + increase_log_steps=True, + rescale=True, + disabled=False, + log_on_batch_idx=False, + log_first_step=False, + log_images_kwargs=None, + log_before_first_step=False, + enable_autocast=True, + ): + super().__init__() + self.enable_autocast = enable_autocast + self.rescale = rescale + self.batch_freq = batch_frequency + self.max_images = max_images + self.log_steps = [2**n for n in range(int(np.log2(self.batch_freq)) + 1)] + if not increase_log_steps: + self.log_steps = [self.batch_freq] + self.clamp = clamp + self.disabled = disabled + self.log_on_batch_idx = log_on_batch_idx + self.log_images_kwargs = log_images_kwargs if log_images_kwargs else {} + self.log_first_step = log_first_step + self.log_before_first_step = log_before_first_step + + @rank_zero_only + def log_local( + self, + save_dir, + split, + images, + global_step, + current_epoch, + batch_idx, + pl_module: Union[None, pl.LightningModule] = None, + ): + root = os.path.join(save_dir, "images", split) + for k in images: + if isheatmap(images[k]): + fig, ax = plt.subplots() + ax = ax.matshow( + images[k].cpu().numpy(), cmap="hot", interpolation="lanczos" + ) + plt.colorbar(ax) + plt.axis("off") + + filename = "{}_gs-{:06}_e-{:06}_b-{:06}.png".format( + k, global_step, current_epoch, batch_idx + ) + os.makedirs(root, exist_ok=True) + path = os.path.join(root, filename) + plt.savefig(path) + plt.close() + # TODO: support wandb + else: + grid = torchvision.utils.make_grid(images[k], nrow=4) + if self.rescale: + grid = (grid + 1.0) / 2.0 # -1,1 -> 0,1; c,h,w + grid = grid.transpose(0, 1).transpose(1, 2).squeeze(-1) + grid = grid.numpy() + grid = (grid * 255).astype(np.uint8) + filename = "{}_gs-{:06}_e-{:06}_b-{:06}.png".format( + k, global_step, current_epoch, batch_idx + ) + path = os.path.join(root, filename) + os.makedirs(os.path.split(path)[0], exist_ok=True) + img = Image.fromarray(grid) + img.save(path) + if exists(pl_module): + assert isinstance( + pl_module.logger, WandbLogger + ), "logger_log_image only supports WandbLogger currently" + pl_module.logger.log_image( + key=f"{split}/{k}", + images=[ + img, + ], + step=pl_module.global_step, + ) + + @rank_zero_only + def log_img(self, pl_module, batch, batch_idx, split="train"): + check_idx = batch_idx if self.log_on_batch_idx else pl_module.global_step + if ( + self.check_frequency(check_idx) + and hasattr(pl_module, "log_images") # batch_idx % self.batch_freq == 0 + and callable(pl_module.log_images) + and + # batch_idx > 5 and + self.max_images > 0 + ): + logger = type(pl_module.logger) + is_train = pl_module.training + if is_train: + pl_module.eval() + + gpu_autocast_kwargs = { + "enabled": self.enable_autocast, # torch.is_autocast_enabled(), + "dtype": torch.get_autocast_gpu_dtype(), + "cache_enabled": torch.is_autocast_cache_enabled(), + } + with torch.no_grad(), torch.cuda.amp.autocast(**gpu_autocast_kwargs): + images = pl_module.log_images( + batch, split=split, **self.log_images_kwargs + ) + + for k in images: + N = min(images[k].shape[0], self.max_images) + if not isheatmap(images[k]): + images[k] = images[k][:N] + if isinstance(images[k], torch.Tensor): + images[k] = images[k].detach().float().cpu() + if self.clamp and not isheatmap(images[k]): + images[k] = torch.clamp(images[k], -1.0, 1.0) + + self.log_local( + pl_module.logger.save_dir, + split, + images, + pl_module.global_step, + pl_module.current_epoch, + batch_idx, + pl_module=pl_module + if isinstance(pl_module.logger, WandbLogger) + else None, + ) + + if is_train: + pl_module.train() + + def check_frequency(self, check_idx): + if ((check_idx % self.batch_freq) == 0 or (check_idx in self.log_steps)) and ( + check_idx > 0 or self.log_first_step + ): + try: + self.log_steps.pop(0) + except IndexError as e: + print(e) + pass + return True + return False + + @rank_zero_only + def on_train_batch_end(self, trainer, pl_module, outputs, batch, batch_idx): + if not self.disabled and (pl_module.global_step > 0 or self.log_first_step): + # self.log_img(pl_module, batch, batch_idx, split="train") + pass + + @rank_zero_only + def on_train_batch_start(self, trainer, pl_module, batch, batch_idx): + if self.log_before_first_step and pl_module.global_step == 0: + print(f"{self.__class__.__name__}: logging before training") + # self.log_img(pl_module, batch, batch_idx, split="train") + pass + + @rank_zero_only + def on_validation_batch_end( + self, trainer, pl_module, outputs, batch, batch_idx, *args, **kwargs + ): + if not self.disabled and pl_module.global_step > 0: + # self.log_img(pl_module, batch, batch_idx, split="val") + pass + if hasattr(pl_module, "calibrate_grad_norm"): + if ( + pl_module.calibrate_grad_norm and batch_idx % 25 == 0 + ) and batch_idx > 0: + self.log_gradients(trainer, pl_module, batch_idx=batch_idx) + + +@rank_zero_only +def init_wandb(save_dir, opt, config, group_name, name_str): + print(f"setting WANDB_DIR to {save_dir}") + os.makedirs(save_dir, exist_ok=True) + + os.environ["WANDB_DIR"] = save_dir + if opt.debug: + wandb.init(project=opt.projectname, mode="offline", group=group_name) + else: + wandb.init( + project=opt.projectname, + config=config, + settings=wandb.Settings(code_dir="./sgm"), + group=group_name, + name=name_str, + ) + + +if __name__ == "__main__": + # custom parser to specify config files, train, test and debug mode, + # postfix, resume. + # `--key value` arguments are interpreted as arguments to the trainer. + # `nested.key=value` arguments are interpreted as config parameters. + # configs are merged from left-to-right followed by command line parameters. + + # model: + # base_learning_rate: float + # target: path to lightning module + # params: + # key: value + # data: + # target: main.DataModuleFromConfig + # params: + # batch_size: int + # wrap: bool + # train: + # target: path to train dataset + # params: + # key: value + # validation: + # target: path to validation dataset + # params: + # key: value + # test: + # target: path to test dataset + # params: + # key: value + # lightning: (optional, has sane defaults and can be specified on cmdline) + # trainer: + # additional arguments to trainer + # logger: + # logger to instantiate + # modelcheckpoint: + # modelcheckpoint to instantiate + # callbacks: + # callback1: + # target: importpath + # params: + # key: value + + now = datetime.datetime.now().strftime("%Y-%m-%dT%H-%M-%S") + + # add cwd for convenience and to make classes in this file available when + # running as `python main.py` + # (in particular `main.DataModuleFromConfig`) + sys.path.append(os.getcwd()) + + parser = get_parser() + + opt, unknown = parser.parse_known_args() + + if opt.name and opt.resume: + raise ValueError( + "-n/--name and -r/--resume cannot be specified both." + "If you want to resume training in a new log folder, " + "use -n/--name in combination with --resume_from_checkpoint" + ) + melk_ckpt_name = None + name = None + if opt.resume: + if not os.path.exists(opt.resume): + raise ValueError("Cannot find {}".format(opt.resume)) + if os.path.isfile(opt.resume): + paths = opt.resume.split("/") + # idx = len(paths)-paths[::-1].index("logs")+1 + # logdir = "/".join(paths[:idx]) + logdir = "/".join(paths[:-2]) + ckpt = opt.resume + _, melk_ckpt_name = get_checkpoint_name(logdir) + else: + assert os.path.isdir(opt.resume), opt.resume + logdir = opt.resume.rstrip("/") + ckpt, melk_ckpt_name = get_checkpoint_name(logdir) + + print("#" * 100) + print(f'Resuming from checkpoint "{ckpt}"') + print("#" * 100) + + opt.resume_from_checkpoint = ckpt + base_configs = sorted(glob.glob(os.path.join(logdir, "configs/*.yaml"))) + opt.base = base_configs + opt.base + _tmp = logdir.split("/") + nowname = _tmp[-1] + else: + if opt.name: + name = "_" + opt.name + elif opt.base: + if opt.no_base_name: + name = "" + else: + if opt.legacy_naming: + cfg_fname = os.path.split(opt.base[0])[-1] + cfg_name = os.path.splitext(cfg_fname)[0] + else: + assert "configs" in os.path.split(opt.base[0])[0], os.path.split( + opt.base[0] + )[0] + cfg_path = os.path.split(opt.base[0])[0].split(os.sep)[ + os.path.split(opt.base[0])[0].split(os.sep).index("configs") + + 1 : + ] # cut away the first one (we assert all configs are in "configs") + cfg_name = os.path.splitext(os.path.split(opt.base[0])[-1])[0] + cfg_name = "-".join(cfg_path) + f"-{cfg_name}" + name = "_" + cfg_name + else: + name = "" + if not opt.no_date: + nowname = now + name + opt.postfix + else: + nowname = name + opt.postfix + if nowname.startswith("_"): + nowname = nowname[1:] + logdir = os.path.join(opt.logdir, nowname) + print(f"LOGDIR: {logdir}") + + ckptdir = os.path.join(logdir, "checkpoints") + cfgdir = os.path.join(logdir, "configs") + seed_everything(opt.seed, workers=True) + + # move before model init, in case a torch.compile(...) is called somewhere + if opt.enable_tf32: + # pt_version = version.parse(torch.__version__) + torch.backends.cuda.matmul.allow_tf32 = True + torch.backends.cudnn.allow_tf32 = True + print(f"Enabling TF32 for PyTorch {torch.__version__}") + else: + print(f"Using default TF32 settings for PyTorch {torch.__version__}:") + print( + f"torch.backends.cuda.matmul.allow_tf32={torch.backends.cuda.matmul.allow_tf32}" + ) + print(f"torch.backends.cudnn.allow_tf32={torch.backends.cudnn.allow_tf32}") + + try: + # init and save configs + configs = [OmegaConf.load(cfg) for cfg in opt.base] + cli = OmegaConf.from_dotlist(unknown) + config = OmegaConf.merge(*configs, cli) + lightning_config = config.pop("lightning", OmegaConf.create()) + # merge trainer cli with config + trainer_config = lightning_config.get("trainer", OmegaConf.create()) + + # default to gpu + trainer_config["accelerator"] = "gpu" + # + standard_args = default_trainer_args() + for k in standard_args: + if getattr(opt, k) != standard_args[k]: + trainer_config[k] = getattr(opt, k) + + ckpt_resume_path = opt.resume_from_checkpoint + + if not "devices" in trainer_config and trainer_config["accelerator"] != "gpu": + del trainer_config["accelerator"] + cpu = True + else: + gpuinfo = trainer_config["devices"] + print(f"Running on GPUs {gpuinfo}") + cpu = False + trainer_opt = argparse.Namespace(**trainer_config) + lightning_config.trainer = trainer_config + + # model + model = instantiate_from_config(config.model) + + # trainer and callbacks + trainer_kwargs = dict() + + # default logger configs + default_logger_cfgs = { + "wandb": { + "target": "pytorch_lightning.loggers.WandbLogger", + "params": { + "name": nowname, + # "save_dir": logdir, + "offline": opt.debug, + "id": nowname, + "project": opt.projectname, + "log_model": False, + # "dir": logdir, + }, + }, + "csv": { + "target": "pytorch_lightning.loggers.CSVLogger", + "params": { + "name": "testtube", # hack for sbord fanatics + "save_dir": logdir, + }, + }, + } + default_logger_cfg = default_logger_cfgs["wandb" if opt.wandb else "csv"] + if opt.wandb: + # TODO change once leaving "swiffer" config directory + try: + group_name = nowname.split(now)[-1].split("-")[1] + except: + group_name = nowname + default_logger_cfg["params"]["group"] = group_name + init_wandb( + os.path.join(os.getcwd(), logdir), + opt=opt, + group_name=group_name, + config=config, + name_str=nowname, + ) + if "logger" in lightning_config: + logger_cfg = lightning_config.logger + else: + logger_cfg = OmegaConf.create() + logger_cfg = OmegaConf.merge(default_logger_cfg, logger_cfg) + trainer_kwargs["logger"] = instantiate_from_config(logger_cfg) + + # modelcheckpoint - use TrainResult/EvalResult(checkpoint_on=metric) to + # specify which metric is used to determine best models + default_modelckpt_cfg = { + "target": "pytorch_lightning.callbacks.ModelCheckpoint", + "params": { + "dirpath": ckptdir, + "filename": "{epoch:06}", + "verbose": True, + "save_last": True, + }, + } + if hasattr(model, "monitor"): + print(f"Monitoring {model.monitor} as checkpoint metric.") + default_modelckpt_cfg["params"]["monitor"] = model.monitor + default_modelckpt_cfg["params"]["save_top_k"] = 3 + + if "modelcheckpoint" in lightning_config: + modelckpt_cfg = lightning_config.modelcheckpoint + else: + modelckpt_cfg = OmegaConf.create() + modelckpt_cfg = OmegaConf.merge(default_modelckpt_cfg, modelckpt_cfg) + print(f"Merged modelckpt-cfg: \n{modelckpt_cfg}") + + # https://pytorch-lightning.readthedocs.io/en/stable/extensions/strategy.html + # default to ddp if not further specified + default_strategy_config = {"target": "pytorch_lightning.strategies.DDPStrategy"} + + if "strategy" in lightning_config: + strategy_cfg = lightning_config.strategy + else: + strategy_cfg = OmegaConf.create() + default_strategy_config["params"] = { + "find_unused_parameters": False, + # "static_graph": True, + # "ddp_comm_hook": default.fp16_compress_hook # TODO: experiment with this, also for DDPSharded + } + strategy_cfg = OmegaConf.merge(default_strategy_config, strategy_cfg) + print( + f"strategy config: \n ++++++++++++++ \n {strategy_cfg} \n ++++++++++++++ " + ) + trainer_kwargs["strategy"] = instantiate_from_config(strategy_cfg) + + # add callback which sets up log directory + default_callbacks_cfg = { + "setup_callback": { + "target": "main.SetupCallback", + "params": { + "resume": opt.resume, + "now": now, + "logdir": logdir, + "ckptdir": ckptdir, + "cfgdir": cfgdir, + "config": config, + "lightning_config": lightning_config, + "debug": opt.debug, + "ckpt_name": melk_ckpt_name, + }, + }, + "image_logger": { + "target": "main.ImageLogger", + "params": {"batch_frequency": 1000, "max_images": 4, "clamp": True}, + }, + "learning_rate_logger": { + "target": "pytorch_lightning.callbacks.LearningRateMonitor", + "params": { + "logging_interval": "step", + # "log_momentum": True + }, + }, + } + if version.parse(pl.__version__) >= version.parse("1.4.0"): + default_callbacks_cfg.update({"checkpoint_callback": modelckpt_cfg}) + + if "callbacks" in lightning_config: + callbacks_cfg = lightning_config.callbacks + else: + callbacks_cfg = OmegaConf.create() + + if "metrics_over_trainsteps_checkpoint" in callbacks_cfg: + print( + "Caution: Saving checkpoints every n train steps without deleting. This might require some free space." + ) + default_metrics_over_trainsteps_ckpt_dict = { + "metrics_over_trainsteps_checkpoint": { + "target": "pytorch_lightning.callbacks.ModelCheckpoint", + "params": { + "dirpath": os.path.join(ckptdir, "trainstep_checkpoints"), + "filename": "{epoch:06}-{step:09}", + "verbose": True, + "save_top_k": -1, + "every_n_train_steps": 10000, + "save_weights_only": True, + }, + } + } + default_callbacks_cfg.update(default_metrics_over_trainsteps_ckpt_dict) + + callbacks_cfg = OmegaConf.merge(default_callbacks_cfg, callbacks_cfg) + if "ignore_keys_callback" in callbacks_cfg and ckpt_resume_path is not None: + callbacks_cfg.ignore_keys_callback.params["ckpt_path"] = ckpt_resume_path + elif "ignore_keys_callback" in callbacks_cfg: + del callbacks_cfg["ignore_keys_callback"] + + trainer_kwargs["callbacks"] = [ + instantiate_from_config(callbacks_cfg[k]) for k in callbacks_cfg + ] + if not "plugins" in trainer_kwargs: + trainer_kwargs["plugins"] = list() + + # cmd line trainer args (which are in trainer_opt) have always priority over config-trainer-args (which are in trainer_kwargs) + trainer_opt = vars(trainer_opt) + trainer_kwargs = { + key: val for key, val in trainer_kwargs.items() if key not in trainer_opt + } + trainer = Trainer(**trainer_opt, **trainer_kwargs) + + trainer.logdir = logdir ### + + # data + data = instantiate_from_config(config.data) + # NOTE according to https://pytorch-lightning.readthedocs.io/en/latest/datamodules.html + # calling these ourselves should not be necessary but it is. + # lightning still takes care of proper multiprocessing though + data.prepare_data() + # data.setup() + print("#### Data #####") + try: + for k in data.datasets: + print( + f"{k}, {data.datasets[k].__class__.__name__}, {len(data.datasets[k])}" + ) + except: + print("datasets not yet initialized.") + + # configure learning rate + if "batch_size" in config.data.params: + bs, base_lr = config.data.params.batch_size, config.model.base_learning_rate + else: + bs, base_lr = ( + config.data.params.train.loader.batch_size, + config.model.base_learning_rate, + ) + if not cpu: + ngpu = len(lightning_config.trainer.devices.strip(",").split(",")) + else: + ngpu = 1 + if "accumulate_grad_batches" in lightning_config.trainer: + accumulate_grad_batches = lightning_config.trainer.accumulate_grad_batches + else: + accumulate_grad_batches = 1 + print(f"accumulate_grad_batches = {accumulate_grad_batches}") + lightning_config.trainer.accumulate_grad_batches = accumulate_grad_batches + if opt.scale_lr: + model.learning_rate = accumulate_grad_batches * ngpu * bs * base_lr + print( + "Setting learning rate to {:.2e} = {} (accumulate_grad_batches) * {} (num_gpus) * {} (batchsize) * {:.2e} (base_lr)".format( + model.learning_rate, accumulate_grad_batches, ngpu, bs, base_lr + ) + ) + else: + model.learning_rate = base_lr + print("++++ NOT USING LR SCALING ++++") + print(f"Setting learning rate to {model.learning_rate:.2e}") + + # allow checkpointing via USR1 + def melk(*args, **kwargs): + # run all checkpoint hooks + if trainer.global_rank == 0: + print("Summoning checkpoint.") + if melk_ckpt_name is None: + ckpt_path = os.path.join(ckptdir, "last.ckpt") + else: + ckpt_path = os.path.join(ckptdir, melk_ckpt_name) + trainer.save_checkpoint(ckpt_path) + + def divein(*args, **kwargs): + if trainer.global_rank == 0: + import pudb + + pudb.set_trace() + + import signal + + signal.signal(signal.SIGUSR1, melk) + signal.signal(signal.SIGUSR2, divein) + + # run + if opt.train: + try: + trainer.fit(model, data, ckpt_path=ckpt_resume_path) + except Exception: + if not opt.debug: + melk() + raise + if not opt.no_test and not trainer.interrupted: + trainer.test(model, data) + except RuntimeError as err: + if MULTINODE_HACKS: + import datetime + import os + import socket + + import requests + + device = os.environ.get("CUDA_VISIBLE_DEVICES", "?") + hostname = socket.gethostname() + ts = datetime.datetime.utcnow().strftime("%Y-%m-%d %H:%M:%S") + resp = requests.get("http://169.254.169.254/latest/meta-data/instance-id") + print( + f"ERROR at {ts} on {hostname}/{resp.text} (CUDA_VISIBLE_DEVICES={device}): {type(err).__name__}: {err}", + flush=True, + ) + raise err + except Exception: + if opt.debug and trainer.global_rank == 0: + try: + import pudb as debugger + except ImportError: + import pdb as debugger + debugger.post_mortem() + raise + finally: + # move newly created debug project to debug_runs + if opt.debug and not opt.resume and trainer.global_rank == 0: + dst, name = os.path.split(logdir) + dst = os.path.join(dst, "debug_runs", name) + os.makedirs(os.path.split(dst)[0], exist_ok=True) + os.rename(logdir, dst) + + if opt.wandb: + wandb.finish() + # if trainer.global_rank == 0: + # print(trainer.profiler.summary()) diff --git a/precompute_latents.py b/precompute_latents.py new file mode 100644 index 0000000..c01e728 --- /dev/null +++ b/precompute_latents.py @@ -0,0 +1,68 @@ +import os +import torch +from torchvision import transforms +from PIL import Image +from diffusers import AutoencoderKL +from tqdm import tqdm + +# Load the VAE model of Stable Diffusion 2.1 +vae = AutoencoderKL.from_pretrained("stabilityai/stable-diffusion-2-1-base", subfolder="vae").cuda() +vae.eval() + +# Define the dataset directory and target directory +DATASET_DIR = "/your/path/DL3DV-ALL-960P" +TARGET_DIR = "/your/path/DL3DV-ALL-960P-latents" + +# Define image preprocessing +preprocess = transforms.Compose([ + transforms.CenterCrop(540), # Center crop to square (DL3DV images are 940x540) + transforms.Resize((576, 576)), # Resize to 512x512 + transforms.ToTensor(), # Convert to tensor + transforms.Normalize([0.5], [0.5]) # Normalize to [-1, 1] +]) + +# Function to process images and save latents +def process_and_save_latents(scene_path, target_scene_path): + os.makedirs(target_scene_path, exist_ok=True) + images_dir = os.path.join(scene_path, "images_4") + target_images_dir = os.path.join(target_scene_path, "images_4") + os.makedirs(target_images_dir, exist_ok=True) + + image_names = [name for name in os.listdir(images_dir) if name.lower().endswith(('.png', '.jpg', '.jpeg'))] + batch_size = 32 + + for i in range(0, len(image_names), batch_size): + batch_names = image_names[i:i + batch_size] + batch_images = [] + + # Load and preprocess images in the batch + for image_name in batch_names: + image_path = os.path.join(images_dir, image_name) + image = Image.open(image_path).convert("RGB") + image_tensor = preprocess(image) + batch_images.append(image_tensor) + + # Stack images into a batch tensor + batch_tensor = torch.stack(batch_images).cuda() + + # Encode the batch to latents + with torch.no_grad(): + latents = vae.encode(batch_tensor).latent_dist.sample() + latents = latents.cpu() # Move to CPU + + # Save each latent in the batch + for j, image_name in enumerate(batch_names): + latent_path = os.path.join(target_images_dir, f"{os.path.splitext(image_name)[0]}.pt") + torch.save(latents[j], latent_path) + +# Iterate through the dataset and process each scene +for subfolder_name in tqdm(os.listdir(DATASET_DIR), desc="Processing subfolders"): + subfolder_path = os.path.join(DATASET_DIR, subfolder_name) + if os.path.isdir(subfolder_path): + for scene_name in tqdm(os.listdir(subfolder_path), desc=f"Processing scenes in {subfolder_name}", leave=False): + scene_path = os.path.join(subfolder_path, scene_name) + if os.path.isdir(scene_path): + target_scene_path = os.path.join(TARGET_DIR, subfolder_name, scene_name) + process_and_save_latents(scene_path, target_scene_path) + +print("Processing complete. Latents saved.") \ No newline at end of file diff --git a/seva/model.py b/seva/model.py index c5d719c..35aa06e 100644 --- a/seva/model.py +++ b/seva/model.py @@ -31,6 +31,7 @@ class SevaParams(object): unflatten_names: list[str] = field( default_factory=lambda: ["middle_ds8", "output_ds4", "output_ds2"] ) + ckpt_path: str | None = None def __post_init__(self): assert len(self.channel_mult) == len(self.transformer_depth) @@ -172,6 +173,12 @@ def __init__(self, params: SevaParams) -> None: nn.SiLU(), nn.Conv2d(self.model_channels, params.out_channels, 3, padding=1), ) + if params.ckpt_path is not None: + from safetensors.torch import load_file + from seva.utils import print_load_warning + state_dict = load_file(params.ckpt_path) + missing, unexpected = self.load_state_dict(state_dict, strict=False, assign=True) + print_load_warning(missing, unexpected) def forward( self, diff --git a/sgm/__init__.py b/sgm/__init__.py new file mode 100644 index 0000000..24bc84a --- /dev/null +++ b/sgm/__init__.py @@ -0,0 +1,4 @@ +from .models import AutoencodingEngine, DiffusionEngine +from .util import get_configs_path, instantiate_from_config + +__version__ = "0.1.0" diff --git a/sgm/data/__init__.py b/sgm/data/__init__.py new file mode 100644 index 0000000..9cac0cb --- /dev/null +++ b/sgm/data/__init__.py @@ -0,0 +1 @@ +from .dataset import StableDataModuleFromConfig, DL3DVDataModuleFromConfig diff --git a/sgm/data/cifar10.py b/sgm/data/cifar10.py new file mode 100644 index 0000000..6083646 --- /dev/null +++ b/sgm/data/cifar10.py @@ -0,0 +1,67 @@ +import pytorch_lightning as pl +import torchvision +from torch.utils.data import DataLoader, Dataset +from torchvision import transforms + + +class CIFAR10DataDictWrapper(Dataset): + def __init__(self, dset): + super().__init__() + self.dset = dset + + def __getitem__(self, i): + x, y = self.dset[i] + return {"jpg": x, "cls": y} + + def __len__(self): + return len(self.dset) + + +class CIFAR10Loader(pl.LightningDataModule): + def __init__(self, batch_size, num_workers=0, shuffle=True): + super().__init__() + + transform = transforms.Compose( + [transforms.ToTensor(), transforms.Lambda(lambda x: x * 2.0 - 1.0)] + ) + + self.batch_size = batch_size + self.num_workers = num_workers + self.shuffle = shuffle + self.train_dataset = CIFAR10DataDictWrapper( + torchvision.datasets.CIFAR10( + root=".data/", train=True, download=True, transform=transform + ) + ) + self.test_dataset = CIFAR10DataDictWrapper( + torchvision.datasets.CIFAR10( + root=".data/", train=False, download=True, transform=transform + ) + ) + + def prepare_data(self): + pass + + def train_dataloader(self): + return DataLoader( + self.train_dataset, + batch_size=self.batch_size, + shuffle=self.shuffle, + num_workers=self.num_workers, + ) + + def test_dataloader(self): + return DataLoader( + self.test_dataset, + batch_size=self.batch_size, + shuffle=self.shuffle, + num_workers=self.num_workers, + ) + + def val_dataloader(self): + return DataLoader( + self.test_dataset, + batch_size=self.batch_size, + shuffle=self.shuffle, + num_workers=self.num_workers, + ) diff --git a/sgm/data/dataset.py b/sgm/data/dataset.py new file mode 100644 index 0000000..9163520 --- /dev/null +++ b/sgm/data/dataset.py @@ -0,0 +1,322 @@ +from typing import Optional +import webdataset as wds +from omegaconf import DictConfig +from pytorch_lightning import LightningDataModule + +from torch.utils.data import Dataset +import os +import json +from PIL import Image +import torch +import cv2 as cv +import numpy as np + + +from seva.geometry import get_plucker_coordinates +from sgm.data.read_write_model import read_model +from sgm.data.utils_camera import ( + read_intrinsics_colmap, + read_extrinsics_colmap, + read_intrinsics_nerfstudio, + read_extrinsics_nerfstudio, + opencv_to_opengl, + colmap_to_nerfstudio, + nerfstudio_to_colmap +) +from einops import rearrange, repeat +import webdataset as wds +from torch.utils.data import DataLoader +from torchvision import transforms + +# try: TODO: fix torchdata version +# from sdata import create_dataset, create_dummy_dataset, create_loader +# except ImportError as e: +# print("#" * 100) +# print("Datasets not yet available") +# print("to enable, we need to add stable-datasets as a submodule") +# print("please use ``git submodule update --init --recursive``") +# print("and do ``pip install -e stable-datasets/`` from the root of this repo") +# print("#" * 100) +# exit(1) + + +class StableDataModuleFromConfig(LightningDataModule): + def __init__( + self, + train: DictConfig, + validation: Optional[DictConfig] = None, + test: Optional[DictConfig] = None, + skip_val_loader: bool = False, + dummy: bool = False, + ): + super().__init__() + self.train_config = train + assert ( + "datapipeline" in self.train_config and "loader" in self.train_config + ), "train config requires the fields `datapipeline` and `loader`" + + self.val_config = validation + if not skip_val_loader: + if self.val_config is not None: + assert ( + "datapipeline" in self.val_config and "loader" in self.val_config + ), "validation config requires the fields `datapipeline` and `loader`" + else: + print( + "Warning: No Validation datapipeline defined, using that one from training" + ) + self.val_config = train + + self.test_config = test + if self.test_config is not None: + assert ( + "datapipeline" in self.test_config and "loader" in self.test_config + ), "test config requires the fields `datapipeline` and `loader`" + + self.dummy = dummy + if self.dummy: + print("#" * 100) + print("USING DUMMY DATASET: HOPE YOU'RE DEBUGGING ;)") + print("#" * 100) + + def setup(self, stage: str) -> None: + print("Preparing datasets") + if self.dummy: + data_fn = create_dummy_dataset + else: + data_fn = create_dataset + + self.train_datapipeline = data_fn(**self.train_config.datapipeline) + if self.val_config: + self.val_datapipeline = data_fn(**self.val_config.datapipeline) + if self.test_config: + self.test_datapipeline = data_fn(**self.test_config.datapipeline) + + def train_dataloader(self): + loader = create_loader(self.train_datapipeline, **self.train_config.loader) + return loader + + def val_dataloader(self) -> wds.DataPipeline: + return create_loader(self.val_datapipeline, **self.val_config.loader) + + def test_dataloader(self) -> wds.DataPipeline: + return create_loader(self.test_datapipeline, **self.test_config.loader) + + +def center_cameras(all_c2ws, c2ws): + ref_c2ws = all_c2ws + camera_dist_2med = torch.norm( + ref_c2ws[:, :3, 3] - ref_c2ws[:, :3, 3].median(0, keepdim=True).values, + dim=-1, + ) + valid_mask = camera_dist_2med <= torch.clamp( + torch.quantile(camera_dist_2med, 0.97) * 10, + max=1e6, + ) + c2ws[:, :3, 3] -= ref_c2ws[valid_mask, :3, 3].mean(0, keepdim=True) + + +def scale_cameras(c2ws, camera_scale=2.0): + camera_dists = c2ws[:, :3, 3].clone() + translation_scaling_factor = ( + camera_scale + if torch.isclose( + torch.norm(camera_dists[0]), + torch.zeros(1), + atol=1e-5, + ).any() + else (camera_scale / torch.norm(camera_dists[0])) + ) + c2ws[:, :3, 3] *= translation_scaling_factor + + +class DL3DVDataset(Dataset): + def __init__(self, dataset_dir, colmap_dir, num_images, latents_dir=None, transform=None, levels=None): + self.dataset_dir = dataset_dir + self.colmap_dir = colmap_dir + self.latents_dir = latents_dir + self.transform = transform + self.levels = levels if levels else [] + self.num_images = num_images + self.scenes = self._load_scenes() + self.adjacent_frame_sampling_prob = 0.2 + + if "480P" in dataset_dir: + self.image_shape = (270, 480) + self.images_folder = "images_8" + elif "960P" in dataset_dir: + self.image_shape = (540, 960) + self.images_folder = "images_4" + + self.target_shape = (576, 576) + self.transform = transforms.Compose([ + transforms.CenterCrop(self.image_shape[0]), # Center crop to square + transforms.Resize(self.target_shape), # Resize to target shape + transforms.ToTensor(), # Convert to tensor + transforms.Normalize([0.5], [0.5]) # Normalize to [-1, 1] + ]) + + # Values fro SD 2.1 autoencoder + self.donwsample_factor = 8 + self.scale_factor = 0.18215 + + def _load_scenes(self): + scenes = [] + for level in os.listdir(self.dataset_dir): + if self.levels and level not in self.levels: + continue + level_path = os.path.join(self.dataset_dir, level) + if os.path.isdir(level_path): + for scene in os.listdir(level_path): + scene_path = os.path.join(level_path, scene) + if os.path.isdir(scene_path): + scenes.append(scene_path) + return scenes + + def __len__(self): + return len(self.scenes) + + def __getitem__(self, idx): + scene_path = self.scenes[idx] + + # Load colmap data + colmap_scene_path = os.path.join( + self.colmap_dir, os.path.relpath(scene_path, self.dataset_dir), "colmap", "sparse", "0" + ) + cameras_metas, images_metas, _ = read_model(colmap_scene_path) + images_metas = list(sorted(images_metas.values(), key=lambda x: x.name)) + + # Load images files + images_dir = os.path.join(scene_path, self.images_folder) + images_files = [os.path.join(images_dir, image.name) for image in images_metas] + + # Sample frames indices + if np.random.rand() <= self.adjacent_frame_sampling_prob: + start_idx = np.random.randint(0, len(images_files) - self.num_images) + images_idxs = np.arange(start_idx, start_idx + self.num_images) + else: + images_idxs = np.random.choice(len(images_files), self.num_images, replace=False) + + images_files = [images_files[i] for i in images_idxs] + + clean_latents = None + if self.latents_dir is not None: + # Load latents + latents_dir = images_dir.replace(self.dataset_dir, self.latents_dir) + latents_files = sorted([f for f in os.listdir(latents_dir) if f.endswith(".pt")]) + latents_files = [latents_files[i] for i in images_idxs] + clean_latents = torch.stack([torch.load(os.path.join(latents_dir, latents_files[i])) for i in range(len(latents_files))]) + + # Load frames + frames = torch.zeros((self.num_images, 3, self.target_shape[0], self.target_shape[1])) + for i, img_file in enumerate(images_files): + img_path = os.path.join(images_dir, img_file) + image = Image.open(img_path).convert("RGB") + image = self.transform(image) + frames[i] = image + + # Read extrinsics from COLMAP + all_c2ws = np.array([read_extrinsics_colmap(image_meta, mode="c2w") for image_meta in images_metas]) + all_c2ws = torch.from_numpy(all_c2ws).float() + c2ws = all_c2ws[images_idxs] + center_cameras(all_c2ws, c2ws) + scale_cameras(c2ws) + + # Read intrinsics from COLMAP + intrinsics = read_intrinsics_colmap(cameras_metas[1], normalize=True) + Ks = repeat(intrinsics, 'd1 d2 -> n d1 d2', n=self.num_images) + Ks = torch.from_numpy(Ks).float() + + # Sample input and target frames + num_input_frames = np.random.randint(1, self.num_images) # Randomly select number of input frames + input_frames_indices = np.random.choice(self.num_images, num_input_frames, replace=False) # Randomly select the input frames + + # Create masks + input_frames_mask = torch.zeros(self.num_images, dtype=torch.bool) + input_frames_mask[input_frames_indices] = True + + camera_mask = torch.ones(self.num_images, dtype=torch.bool) + + w2cs = torch.linalg.inv(c2ws) + pluckers = get_plucker_coordinates( + extrinsics_src=w2cs[input_frames_indices[0]], + extrinsics=w2cs, + intrinsics=Ks.clone(), + target_size=(self.target_shape[0] // self.donwsample_factor, + self.target_shape[1] // self.donwsample_factor), + ) + + concat = torch.cat( # binary mask and plcukers + [ + repeat( + input_frames_mask, + "n -> n 1 h w", + h=pluckers.shape[2], + w=pluckers.shape[3], + ), + pluckers, + ], + dim=1, + ) + + replace = torch.cat( # clean latents and binary mask + [ + clean_latents * self.scale_factor, + repeat( + input_frames_mask, + "n -> n 1 h w", + h=pluckers.shape[2], + w=pluckers.shape[3], + ), + ], + dim=1, + ) + + output_dict = { + "clean_latent": clean_latents, + "mask": input_frames_mask, + "plucker": pluckers, + "camera_mask": camera_mask, + "concat": concat, + "frames": frames, + "replace": replace, + } + + return output_dict + + +class DL3DVDataModuleFromConfig(LightningDataModule): + def __init__( + self, + dataset_dir, + colmap_dir, + batch_size, + latents_dir=None, + num_workers=0, + num_images=21, + shuffle=True): + super().__init__() + + self.dataset_dir = dataset_dir + self.colmap_dir = colmap_dir + self.latents_dir = latents_dir + self.batch_size = batch_size + self.num_workers = num_workers + self.shuffle = shuffle + self.train_dataset = DL3DVDataset( + dataset_dir, + colmap_dir, + latents_dir=latents_dir, + num_images=num_images, + ) + + def prepare_data(self): + pass + + def train_dataloader(self): + return DataLoader( + self.train_dataset, + batch_size=self.batch_size, + shuffle=self.shuffle, + num_workers=self.num_workers, + ) diff --git a/sgm/data/mnist.py b/sgm/data/mnist.py new file mode 100644 index 0000000..dea4d7e --- /dev/null +++ b/sgm/data/mnist.py @@ -0,0 +1,85 @@ +import pytorch_lightning as pl +import torchvision +from torch.utils.data import DataLoader, Dataset +from torchvision import transforms + + +class MNISTDataDictWrapper(Dataset): + def __init__(self, dset): + super().__init__() + self.dset = dset + + def __getitem__(self, i): + x, y = self.dset[i] + return {"jpg": x, "cls": y} + + def __len__(self): + return len(self.dset) + + +class MNISTLoader(pl.LightningDataModule): + def __init__(self, batch_size, num_workers=0, prefetch_factor=2, shuffle=True): + super().__init__() + + transform = transforms.Compose( + [transforms.ToTensor(), transforms.Lambda(lambda x: x * 2.0 - 1.0)] + ) + + self.batch_size = batch_size + self.num_workers = num_workers + self.prefetch_factor = prefetch_factor if num_workers > 0 else 0 + self.shuffle = shuffle + self.train_dataset = MNISTDataDictWrapper( + torchvision.datasets.MNIST( + root=".data/", train=True, download=True, transform=transform + ) + ) + self.test_dataset = MNISTDataDictWrapper( + torchvision.datasets.MNIST( + root=".data/", train=False, download=True, transform=transform + ) + ) + + def prepare_data(self): + pass + + def train_dataloader(self): + return DataLoader( + self.train_dataset, + batch_size=self.batch_size, + shuffle=self.shuffle, + num_workers=self.num_workers, + prefetch_factor=self.prefetch_factor, + ) + + def test_dataloader(self): + return DataLoader( + self.test_dataset, + batch_size=self.batch_size, + shuffle=self.shuffle, + num_workers=self.num_workers, + prefetch_factor=self.prefetch_factor, + ) + + def val_dataloader(self): + return DataLoader( + self.test_dataset, + batch_size=self.batch_size, + shuffle=self.shuffle, + num_workers=self.num_workers, + prefetch_factor=self.prefetch_factor, + ) + + +if __name__ == "__main__": + dset = MNISTDataDictWrapper( + torchvision.datasets.MNIST( + root=".data/", + train=False, + download=True, + transform=transforms.Compose( + [transforms.ToTensor(), transforms.Lambda(lambda x: x * 2.0 - 1.0)] + ), + ) + ) + ex = dset[0] diff --git a/sgm/data/read_write_model.py b/sgm/data/read_write_model.py new file mode 100644 index 0000000..beb60fe --- /dev/null +++ b/sgm/data/read_write_model.py @@ -0,0 +1,605 @@ +# Copyright (c) 2023, ETH Zurich and UNC Chapel Hill. +# All rights reserved. +# +# Redistribution and use in source and binary forms, with or without +# modification, are permitted provided that the following conditions are met: +# +# * Redistributions of source code must retain the above copyright +# notice, this list of conditions and the following disclaimer. +# +# * Redistributions in binary form must reproduce the above copyright +# notice, this list of conditions and the following disclaimer in the +# documentation and/or other materials provided with the distribution. +# +# * Neither the name of ETH Zurich and UNC Chapel Hill nor the names of +# its contributors may be used to endorse or promote products derived +# from this software without specific prior written permission. +# +# THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" +# AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE +# IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE +# ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDERS OR CONTRIBUTORS BE +# LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR +# CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF +# SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS +# INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN +# CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) +# ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE +# POSSIBILITY OF SUCH DAMAGE. + + +import os +import collections +import numpy as np +import struct +import argparse + + +CameraModel = collections.namedtuple( + "CameraModel", ["model_id", "model_name", "num_params"] +) +Camera = collections.namedtuple( + "Camera", ["id", "model", "width", "height", "params"] +) +BaseImage = collections.namedtuple( + "Image", ["id", "qvec", "tvec", "camera_id", "name", "xys", "point3D_ids"] +) +Point3D = collections.namedtuple( + "Point3D", ["id", "xyz", "rgb", "error", "image_ids", "point2D_idxs"] +) + + +class Image(BaseImage): + def qvec2rotmat(self): + return qvec2rotmat(self.qvec) + + +CAMERA_MODELS = { + CameraModel(model_id=0, model_name="SIMPLE_PINHOLE", num_params=3), + CameraModel(model_id=1, model_name="PINHOLE", num_params=4), + CameraModel(model_id=2, model_name="SIMPLE_RADIAL", num_params=4), + CameraModel(model_id=3, model_name="RADIAL", num_params=5), + CameraModel(model_id=4, model_name="OPENCV", num_params=8), + CameraModel(model_id=5, model_name="OPENCV_FISHEYE", num_params=8), + CameraModel(model_id=6, model_name="FULL_OPENCV", num_params=12), + CameraModel(model_id=7, model_name="FOV", num_params=5), + CameraModel(model_id=8, model_name="SIMPLE_RADIAL_FISHEYE", num_params=4), + CameraModel(model_id=9, model_name="RADIAL_FISHEYE", num_params=5), + CameraModel(model_id=10, model_name="THIN_PRISM_FISHEYE", num_params=12), +} +CAMERA_MODEL_IDS = dict( + [(camera_model.model_id, camera_model) for camera_model in CAMERA_MODELS] +) +CAMERA_MODEL_NAMES = dict( + [(camera_model.model_name, camera_model) for camera_model in CAMERA_MODELS] +) + + +def read_next_bytes(fid, num_bytes, format_char_sequence, endian_character="<"): + """Read and unpack the next bytes from a binary file. + :param fid: + :param num_bytes: Sum of combination of {2, 4, 8}, e.g. 2, 6, 16, 30, etc. + :param format_char_sequence: List of {c, e, f, d, h, H, i, I, l, L, q, Q}. + :param endian_character: Any of {@, =, <, >, !} + :return: Tuple of read and unpacked values. + """ + data = fid.read(num_bytes) + return struct.unpack(endian_character + format_char_sequence, data) + + +def write_next_bytes(fid, data, format_char_sequence, endian_character="<"): + """pack and write to a binary file. + :param fid: + :param data: data to send, if multiple elements are sent at the same time, + they should be encapsuled either in a list or a tuple + :param format_char_sequence: List of {c, e, f, d, h, H, i, I, l, L, q, Q}. + should be the same length as the data list or tuple + :param endian_character: Any of {@, =, <, >, !} + """ + if isinstance(data, (list, tuple)): + bytes = struct.pack(endian_character + format_char_sequence, *data) + else: + bytes = struct.pack(endian_character + format_char_sequence, data) + fid.write(bytes) + + +def read_cameras_text(path): + """ + see: src/colmap/scene/reconstruction.cc + void Reconstruction::WriteCamerasText(const std::string& path) + void Reconstruction::ReadCamerasText(const std::string& path) + """ + cameras = {} + with open(path, "r") as fid: + while True: + line = fid.readline() + if not line: + break + line = line.strip() + if len(line) > 0 and line[0] != "#": + elems = line.split() + camera_id = int(elems[0]) + model = elems[1] + width = int(elems[2]) + height = int(elems[3]) + params = np.array(tuple(map(float, elems[4:]))) + cameras[camera_id] = Camera( + id=camera_id, + model=model, + width=width, + height=height, + params=params, + ) + return cameras + + +def read_cameras_binary(path_to_model_file): + """ + see: src/colmap/scene/reconstruction.cc + void Reconstruction::WriteCamerasBinary(const std::string& path) + void Reconstruction::ReadCamerasBinary(const std::string& path) + """ + cameras = {} + with open(path_to_model_file, "rb") as fid: + num_cameras = read_next_bytes(fid, 8, "Q")[0] + for _ in range(num_cameras): + camera_properties = read_next_bytes( + fid, num_bytes=24, format_char_sequence="iiQQ" + ) + camera_id = camera_properties[0] + model_id = camera_properties[1] + model_name = CAMERA_MODEL_IDS[camera_properties[1]].model_name + width = camera_properties[2] + height = camera_properties[3] + num_params = CAMERA_MODEL_IDS[model_id].num_params + params = read_next_bytes( + fid, + num_bytes=8 * num_params, + format_char_sequence="d" * num_params, + ) + cameras[camera_id] = Camera( + id=camera_id, + model=model_name, + width=width, + height=height, + params=np.array(params), + ) + assert len(cameras) == num_cameras + return cameras + + +def write_cameras_text(cameras, path): + """ + see: src/colmap/scene/reconstruction.cc + void Reconstruction::WriteCamerasText(const std::string& path) + void Reconstruction::ReadCamerasText(const std::string& path) + """ + HEADER = ( + "# Camera list with one line of data per camera:\n" + + "# CAMERA_ID, MODEL, WIDTH, HEIGHT, PARAMS[]\n" + + "# Number of cameras: {}\n".format(len(cameras)) + ) + with open(path, "w") as fid: + fid.write(HEADER) + for _, cam in cameras.items(): + to_write = [cam.id, cam.model, cam.width, cam.height, *cam.params] + line = " ".join([str(elem) for elem in to_write]) + fid.write(line + "\n") + + +def write_cameras_binary(cameras, path_to_model_file): + """ + see: src/colmap/scene/reconstruction.cc + void Reconstruction::WriteCamerasBinary(const std::string& path) + void Reconstruction::ReadCamerasBinary(const std::string& path) + """ + with open(path_to_model_file, "wb") as fid: + write_next_bytes(fid, len(cameras), "Q") + for _, cam in cameras.items(): + model_id = CAMERA_MODEL_NAMES[cam.model].model_id + camera_properties = [cam.id, model_id, cam.width, cam.height] + write_next_bytes(fid, camera_properties, "iiQQ") + for p in cam.params: + write_next_bytes(fid, float(p), "d") + return cameras + + +def read_images_text(path): + """ + see: src/colmap/scene/reconstruction.cc + void Reconstruction::ReadImagesText(const std::string& path) + void Reconstruction::WriteImagesText(const std::string& path) + """ + images = {} + with open(path, "r") as fid: + while True: + line = fid.readline() + if not line: + break + line = line.strip() + if len(line) > 0 and line[0] != "#": + elems = line.split() + image_id = int(elems[0]) + qvec = np.array(tuple(map(float, elems[1:5]))) + tvec = np.array(tuple(map(float, elems[5:8]))) + camera_id = int(elems[8]) + image_name = elems[9] + elems = fid.readline().split() + xys = np.column_stack( + [ + tuple(map(float, elems[0::3])), + tuple(map(float, elems[1::3])), + ] + ) + point3D_ids = np.array(tuple(map(int, elems[2::3]))) + images[image_id] = Image( + id=image_id, + qvec=qvec, + tvec=tvec, + camera_id=camera_id, + name=image_name, + xys=xys, + point3D_ids=point3D_ids, + ) + return images + + +def read_images_binary(path_to_model_file): + """ + see: src/colmap/scene/reconstruction.cc + void Reconstruction::ReadImagesBinary(const std::string& path) + void Reconstruction::WriteImagesBinary(const std::string& path) + """ + images = {} + with open(path_to_model_file, "rb") as fid: + num_reg_images = read_next_bytes(fid, 8, "Q")[0] + for _ in range(num_reg_images): + binary_image_properties = read_next_bytes( + fid, num_bytes=64, format_char_sequence="idddddddi" + ) + image_id = binary_image_properties[0] + qvec = np.array(binary_image_properties[1:5]) + tvec = np.array(binary_image_properties[5:8]) + camera_id = binary_image_properties[8] + image_name = "" + current_char = read_next_bytes(fid, 1, "c")[0] + while current_char != b"\x00": # look for the ASCII 0 entry + image_name += current_char.decode("utf-8") + current_char = read_next_bytes(fid, 1, "c")[0] + num_points2D = read_next_bytes( + fid, num_bytes=8, format_char_sequence="Q" + )[0] + x_y_id_s = read_next_bytes( + fid, + num_bytes=24 * num_points2D, + format_char_sequence="ddq" * num_points2D, + ) + xys = np.column_stack( + [ + tuple(map(float, x_y_id_s[0::3])), + tuple(map(float, x_y_id_s[1::3])), + ] + ) + point3D_ids = np.array(tuple(map(int, x_y_id_s[2::3]))) + images[image_id] = Image( + id=image_id, + qvec=qvec, + tvec=tvec, + camera_id=camera_id, + name=image_name, + xys=xys, + point3D_ids=point3D_ids, + ) + return images + + +def write_images_text(images, path): + """ + see: src/colmap/scene/reconstruction.cc + void Reconstruction::ReadImagesText(const std::string& path) + void Reconstruction::WriteImagesText(const std::string& path) + """ + if len(images) == 0: + mean_observations = 0 + else: + mean_observations = sum( + (len(img.point3D_ids) for _, img in images.items()) + ) / len(images) + HEADER = ( + "# Image list with two lines of data per image:\n" + + "# IMAGE_ID, QW, QX, QY, QZ, TX, TY, TZ, CAMERA_ID, NAME\n" + + "# POINTS2D[] as (X, Y, POINT3D_ID)\n" + + "# Number of images: {}, mean observations per image: {}\n".format( + len(images), mean_observations + ) + ) + + with open(path, "w") as fid: + fid.write(HEADER) + for _, img in images.items(): + image_header = [ + img.id, + *img.qvec, + *img.tvec, + img.camera_id, + img.name, + ] + first_line = " ".join(map(str, image_header)) + fid.write(first_line + "\n") + + points_strings = [] + for xy, point3D_id in zip(img.xys, img.point3D_ids): + points_strings.append(" ".join(map(str, [*xy, point3D_id]))) + fid.write(" ".join(points_strings) + "\n") + + +def write_images_binary(images, path_to_model_file): + """ + see: src/colmap/scene/reconstruction.cc + void Reconstruction::ReadImagesBinary(const std::string& path) + void Reconstruction::WriteImagesBinary(const std::string& path) + """ + with open(path_to_model_file, "wb") as fid: + write_next_bytes(fid, len(images), "Q") + for _, img in images.items(): + write_next_bytes(fid, img.id, "i") + write_next_bytes(fid, img.qvec.tolist(), "dddd") + write_next_bytes(fid, img.tvec.tolist(), "ddd") + write_next_bytes(fid, img.camera_id, "i") + for char in img.name: + write_next_bytes(fid, char.encode("utf-8"), "c") + write_next_bytes(fid, b"\x00", "c") + write_next_bytes(fid, len(img.point3D_ids), "Q") + for xy, p3d_id in zip(img.xys, img.point3D_ids): + write_next_bytes(fid, [*xy, p3d_id], "ddq") + + +def read_points3D_text(path): + """ + see: src/colmap/scene/reconstruction.cc + void Reconstruction::ReadPoints3DText(const std::string& path) + void Reconstruction::WritePoints3DText(const std::string& path) + """ + points3D = {} + with open(path, "r") as fid: + while True: + line = fid.readline() + if not line: + break + line = line.strip() + if len(line) > 0 and line[0] != "#": + elems = line.split() + point3D_id = int(elems[0]) + xyz = np.array(tuple(map(float, elems[1:4]))) + rgb = np.array(tuple(map(int, elems[4:7]))) + error = float(elems[7]) + image_ids = np.array(tuple(map(int, elems[8::2]))) + point2D_idxs = np.array(tuple(map(int, elems[9::2]))) + points3D[point3D_id] = Point3D( + id=point3D_id, + xyz=xyz, + rgb=rgb, + error=error, + image_ids=image_ids, + point2D_idxs=point2D_idxs, + ) + return points3D + + +def read_points3D_binary(path_to_model_file): + """ + see: src/colmap/scene/reconstruction.cc + void Reconstruction::ReadPoints3DBinary(const std::string& path) + void Reconstruction::WritePoints3DBinary(const std::string& path) + """ + points3D = {} + with open(path_to_model_file, "rb") as fid: + num_points = read_next_bytes(fid, 8, "Q")[0] + for _ in range(num_points): + binary_point_line_properties = read_next_bytes( + fid, num_bytes=43, format_char_sequence="QdddBBBd" + ) + point3D_id = binary_point_line_properties[0] + xyz = np.array(binary_point_line_properties[1:4]) + rgb = np.array(binary_point_line_properties[4:7]) + error = np.array(binary_point_line_properties[7]) + track_length = read_next_bytes( + fid, num_bytes=8, format_char_sequence="Q" + )[0] + track_elems = read_next_bytes( + fid, + num_bytes=8 * track_length, + format_char_sequence="ii" * track_length, + ) + image_ids = np.array(tuple(map(int, track_elems[0::2]))) + point2D_idxs = np.array(tuple(map(int, track_elems[1::2]))) + points3D[point3D_id] = Point3D( + id=point3D_id, + xyz=xyz, + rgb=rgb, + error=error, + image_ids=image_ids, + point2D_idxs=point2D_idxs, + ) + return points3D + + +def write_points3D_text(points3D, path): + """ + see: src/colmap/scene/reconstruction.cc + void Reconstruction::ReadPoints3DText(const std::string& path) + void Reconstruction::WritePoints3DText(const std::string& path) + """ + if len(points3D) == 0: + mean_track_length = 0 + else: + mean_track_length = sum( + (len(pt.image_ids) for _, pt in points3D.items()) + ) / len(points3D) + HEADER = ( + "# 3D point list with one line of data per point:\n" + + "# POINT3D_ID, X, Y, Z, R, G, B, ERROR, TRACK[] as (IMAGE_ID, POINT2D_IDX)\n" + + "# Number of points: {}, mean track length: {}\n".format( + len(points3D), mean_track_length + ) + ) + + with open(path, "w") as fid: + fid.write(HEADER) + for _, pt in points3D.items(): + point_header = [pt.id, *pt.xyz, *pt.rgb, pt.error] + fid.write(" ".join(map(str, point_header)) + " ") + track_strings = [] + for image_id, point2D in zip(pt.image_ids, pt.point2D_idxs): + track_strings.append(" ".join(map(str, [image_id, point2D]))) + fid.write(" ".join(track_strings) + "\n") + + +def write_points3D_binary(points3D, path_to_model_file): + """ + see: src/colmap/scene/reconstruction.cc + void Reconstruction::ReadPoints3DBinary(const std::string& path) + void Reconstruction::WritePoints3DBinary(const std::string& path) + """ + with open(path_to_model_file, "wb") as fid: + write_next_bytes(fid, len(points3D), "Q") + for _, pt in points3D.items(): + write_next_bytes(fid, pt.id, "Q") + write_next_bytes(fid, pt.xyz.tolist(), "ddd") + write_next_bytes(fid, pt.rgb.tolist(), "BBB") + write_next_bytes(fid, pt.error, "d") + track_length = pt.image_ids.shape[0] + write_next_bytes(fid, track_length, "Q") + for image_id, point2D_id in zip(pt.image_ids, pt.point2D_idxs): + write_next_bytes(fid, [image_id, point2D_id], "ii") + + +def detect_model_format(path, ext, verbose=False): + if ( + os.path.isfile(os.path.join(path, "cameras" + ext)) + and os.path.isfile(os.path.join(path, "images" + ext)) + and os.path.isfile(os.path.join(path, "points3D" + ext)) + ): + if verbose: + print("Detected model format: '" + ext + "'") + return True + + return False + + +def read_model(path, ext=""): + # try to detect the extension automatically + if ext == "": + if detect_model_format(path, ".bin"): + ext = ".bin" + elif detect_model_format(path, ".txt"): + ext = ".txt" + else: + print("Provide model format: '.bin' or '.txt'") + return + + if ext == ".txt": + cameras = read_cameras_text(os.path.join(path, "cameras" + ext)) + images = read_images_text(os.path.join(path, "images" + ext)) + points3D = read_points3D_text(os.path.join(path, "points3D") + ext) + else: + cameras = read_cameras_binary(os.path.join(path, "cameras" + ext)) + images = read_images_binary(os.path.join(path, "images" + ext)) + points3D = read_points3D_binary(os.path.join(path, "points3D") + ext) + return cameras, images, points3D + + +def write_model(cameras, images, points3D, path, ext=".bin"): + if ext == ".txt": + write_cameras_text(cameras, os.path.join(path, "cameras" + ext)) + write_images_text(images, os.path.join(path, "images" + ext)) + write_points3D_text(points3D, os.path.join(path, "points3D") + ext) + else: + write_cameras_binary(cameras, os.path.join(path, "cameras" + ext)) + write_images_binary(images, os.path.join(path, "images" + ext)) + write_points3D_binary(points3D, os.path.join(path, "points3D") + ext) + return cameras, images, points3D + + +def qvec2rotmat(qvec): + return np.array( + [ + [ + 1 - 2 * qvec[2] ** 2 - 2 * qvec[3] ** 2, + 2 * qvec[1] * qvec[2] - 2 * qvec[0] * qvec[3], + 2 * qvec[3] * qvec[1] + 2 * qvec[0] * qvec[2], + ], + [ + 2 * qvec[1] * qvec[2] + 2 * qvec[0] * qvec[3], + 1 - 2 * qvec[1] ** 2 - 2 * qvec[3] ** 2, + 2 * qvec[2] * qvec[3] - 2 * qvec[0] * qvec[1], + ], + [ + 2 * qvec[3] * qvec[1] - 2 * qvec[0] * qvec[2], + 2 * qvec[2] * qvec[3] + 2 * qvec[0] * qvec[1], + 1 - 2 * qvec[1] ** 2 - 2 * qvec[2] ** 2, + ], + ] + ) + + +def rotmat2qvec(R): + Rxx, Ryx, Rzx, Rxy, Ryy, Rzy, Rxz, Ryz, Rzz = R.flat + K = ( + np.array( + [ + [Rxx - Ryy - Rzz, 0, 0, 0], + [Ryx + Rxy, Ryy - Rxx - Rzz, 0, 0], + [Rzx + Rxz, Rzy + Ryz, Rzz - Rxx - Ryy, 0], + [Ryz - Rzy, Rzx - Rxz, Rxy - Ryx, Rxx + Ryy + Rzz], + ] + ) + / 3.0 + ) + eigvals, eigvecs = np.linalg.eigh(K) + qvec = eigvecs[[3, 0, 1, 2], np.argmax(eigvals)] + if qvec[0] < 0: + qvec *= -1 + return qvec + + +# def main(): +# parser = argparse.ArgumentParser( +# description="Read and write COLMAP binary and text models" +# ) +# parser.add_argument("--input_model", help="path to input model folder") +# parser.add_argument( +# "--input_format", +# choices=[".bin", ".txt"], +# help="input model format", +# default="", +# ) +# parser.add_argument("--output_model", help="path to output model folder") +# parser.add_argument( +# "--output_format", +# choices=[".bin", ".txt"], +# help="outut model format", +# default=".txt", +# ) +# args = parser.parse_args() + +# cameras, images, points3D = read_model( +# path=args.input_model, ext=args.input_format +# ) + +# print("num_cameras:", len(cameras)) +# print("num_images:", len(images)) +# print("num_points3D:", len(points3D)) + +# if args.output_model is not None: +# write_model( +# cameras, +# images, +# points3D, +# path=args.output_model, +# ext=args.output_format, +# ) + + +# if __name__ == "__main__": +# main() diff --git a/sgm/data/utils_camera.py b/sgm/data/utils_camera.py new file mode 100644 index 0000000..f6ebdbc --- /dev/null +++ b/sgm/data/utils_camera.py @@ -0,0 +1,218 @@ +import os +import numpy as np +import json +import sys +from sgm.data.read_write_model import qvec2rotmat, read_cameras_text, read_images_text + + +def read_extrinsics_colmap(image_meta, mode="c2w"): + """ + Converts image metadata from COLMAP format to a world-to-camera (w2c) transformation matrix. + Args: + image_meta: An object containing the following attributes: + - tvec (numpy.ndarray): A 3-element translation vector. + - qvec (numpy.ndarray): A 4-element quaternion representing rotation. + Returns: + numpy.ndarray: A 4x4 world-to-camera transformation matrix. + """ + + translation = image_meta.tvec.reshape(3, 1) + rotation = qvec2rotmat(image_meta.qvec) + w2c = np.concatenate([rotation, translation], 1) + w2c = np.concatenate([w2c, np.array([[0.0, 0.0, 0.0, 1.0]])], 0) + if mode == "w2c": + return w2c + elif mode == "c2w": + return np.linalg.inv(w2c) + else: + raise ValueError + +def read_intrinsics_colmap(camera_meta, normalize=False): + assert camera_meta.model in {"PINHOLE", "OPENCV"} + K = np.eye(3, dtype=np.float32) + K[0,0] = camera_meta.params[0] # f_x + K[1,1] = camera_meta.params[1] # f_y + K[0,2] = camera_meta.params[2] # c_x + K[1,2] = camera_meta.params[3] # c_y + + if normalize: + K[0,0] /= camera_meta.width + K[1,1] /= camera_meta.height + K[0,2] /= camera_meta.width + K[1,2] /= camera_meta.height + return K + +def read_extrinsics_nerfstudio(transforms_dict, mode="c2w"): + """ + Reads camera extrinsics from a dictionary and returns a list of all 4x4 transformation matrices. + + Args: + transforms_dict (dict): A dictionary containing camera extrinsics with the following keys: + - "frames" (list): A list of dictionaries, each containing: + - "transform_matrix" (list): A 4x4 transformation matrix. + + Returns: + list: A list of numpy.ndarray objects, each representing a 4x4 transformation matrix. + """ + num_frames = len(transforms_dict["frames"]) + c2ws = np.empty((num_frames, 4, 4), dtype=np.float32) + for i, frame in enumerate(transforms_dict["frames"]): + c2ws[i] = np.array(frame["transform_matrix"], dtype=np.float32) + return c2ws + + +def read_intrinsics_nerfstudio(transforms_dict, normalize=False): + """ + Reads camera intrinsics from a dictionary and returns the intrinsic matrix. + + transforms_dict (dict): A dictionary containing camera intrinsics with the following keys: + - "w" (int): Image width. + - "h" (int): Image height. + - "fl_x" (float): Focal length in the x direction. + - "fl_y" (float): Focal length in the y direction. + - "cx" (float): Principal point x-coordinate. + - "cy" (float): Principal point y-coordinate. + principal point coordinates by the image width and height. Defaults to False. + + Returns: + numpy.ndarray: A 3x3 camera intrinsic matrix (K) with the following structure: + [[fl_x, 0, cx], + [ 0, fl_y, cy], + [ 0, 0, 1]] + If `normalize` is True, `fl_x`, `fl_y`, `cx`, and `cy` are normalized + by the image width and height. + """ + + w = transforms_dict["w"] + h = transforms_dict["h"] + fl_x = transforms_dict["fl_x"] + fl_y = transforms_dict["fl_y"] + cx = transforms_dict["cx"] + cy = transforms_dict["cy"] + + if normalize: + fl_x /= float(w) + fl_y /= float(h) + cx /= float(w) + cy /= float(h) + + K= np.array([ + [fl_x, 0, cx], + [0, fl_y, cy], + [0, 0, 1] + ], dtype=np.float32) + return K + + +def subsample_images_metas(images_metas, n): + """ + Subsamples a dictionary of image metadata by selecting `n` evenly spaced entries. + + Args: + images_metas (dict): A dictionary where keys represent metadata identifiers + and values represent the corresponding metadata. + n (int): The number of evenly spaced entries to select from the dictionary. + + Returns: + dict: A new dictionary containing `n` evenly spaced entries from the input dictionary. + """ + indices = np.linspace(0, len(images_metas) - 1, n, dtype=int) + sorted_keys = sorted(images_metas.keys()) + return {key: images_metas[key] for i, key in enumerate(sorted_keys) if i in indices} + + +def opencv_to_opengl(c2w): + """ + Converts a camera-to-world transformation matrix (or a batch of matrices) from OpenCV convention to OpenGL convention. + In the OpenCV convention: + - The camera-to-world matrix's rotation is represented as: + `R = [ u_1 | u_2 | u_3 ]` + where `u_1`, `u_2`, and `u_3` are the column vectors of the matrix. + + - In the OpenGL convention, the camera-to-world matrix's rotation is represented as: + `R = [ v_1 | v_2 | v_3 ]` + where `v_1`, `v_2`, and `v_3` are the column vectors of the matrix. + + The conversion is performed as follows: + - `v_1 = u_1` + - `v_2 = -u_2` + - `v_3 = -u_3` + + Note: + - The translation vector (last column of the matrix) remains unchanged, as it represents the position of the camera in world coordinates. + Args: + c2w (np.ndarray or torch.Tensor): A 4x4 camera-to-world transformation matrix or a batch of Nx4x4 matrices. + - If the input is a single 4x4 matrix, it is treated as a single transformation. + - If the input is a batch of matrices, it should have the shape (N, 4, 4), where N is the batch size. + Returns: + (np.ndarray or torch.Tensor): The transformed camera-to-world matrix (or batch of matrices) in OpenGL convention. + - The shape of the output matches the shape of the input (4x4 or Nx4x4). + + See: + - https://www.songho.ca/opengl/gl_camera.html + - https://colmap.github.io/format.html + """ + dummy_batch = (c2w.ndim == 2) + if dummy_batch: + c2w = c2w[None, ...] + + assert c2w.ndim == 3 and c2w.shape[1:] == (4, 4), "Input must have shape Nx4x4" + + c2w[:, :3, 1:3] *= -1 + if dummy_batch: + c2w = c2w[0] + return c2w + + +def nerfstudio_to_colmap(c2w, applied_transform): + """ + applied_transform comes from the transforms.json + """ + applied_transform = np.concatenate([applied_transform, [[0, 0, 0, 1]]]) + dummy_batch = (c2w.ndim == 2) + if dummy_batch: + c2w = c2w[None, ...] + + assert c2w.ndim == 3 and c2w.shape[1:] == (4, 4), "Input must have shape Nx4x4" + + c2w = applied_transform @ c2w # swap x/y coordinates (or y/z) and negate z is the same operation + c2w = opencv_to_opengl(c2w) # negating second and third element of the basis is the same operation + + if dummy_batch: + c2w = c2w[0] + return c2w + + +def colmap_to_nerfstudio(c2w): + ''' + See: + - https://github.com/nerfstudio-project/nerfstudio/blob/main/nerfstudio/process_data/colmap_utils.py#L409 + - https://github.com/nerfstudio-project/nerfstudio/commit/25a00a09abed042dbe9b7d5078ec30efec6fd7e0 + - https://github.com/nerfstudio-project/nerfstudio/issues/1504?utm_source=chatgpt.com + + This code matches the DL3DV transforms.json + ''' + dummy_batch = (c2w.ndim == 2) + if dummy_batch: + c2w = c2w[None, ...] + + assert c2w.ndim == 3 and c2w.shape[1:] == (4, 4), "Input must have shape Nx4x4" + + c2w = opencv_to_opengl(c2w) + + # Match Nerfstudio viewer convention (+z) from COLMAP (-y) + # TODO: what if the applied transform is np.array([0, 2, 1, 3]) + c2w[:] = c2w[:, np.array([1, 0, 2, 3]), :] # x <-> y + c2w[:, 2, :] *= -1 # z -> -z + + if dummy_batch: + c2w = c2w[0] + return c2w + +if __name__ == "__main__": + colmap_dir="/home/nviolant/TrajectoryCrafter/test/colmap_traj" + cameras = read_cameras_text(os.path.join(colmap_dir, "cameras.txt")) + images_metas = read_images_text(os.path.join(colmap_dir, "images.txt")) + + print(type(cameras[1])) + print(read_intrinsics_colmap(cameras[1])) \ No newline at end of file diff --git a/sgm/inference/api.py b/sgm/inference/api.py new file mode 100644 index 0000000..a359a67 --- /dev/null +++ b/sgm/inference/api.py @@ -0,0 +1,385 @@ +import pathlib +from dataclasses import asdict, dataclass +from enum import Enum +from typing import Optional + +from omegaconf import OmegaConf + +from sgm.inference.helpers import (Img2ImgDiscretizationWrapper, do_img2img, + do_sample) +from sgm.modules.diffusionmodules.sampling import (DPMPP2MSampler, + DPMPP2SAncestralSampler, + EulerAncestralSampler, + EulerEDMSampler, + HeunEDMSampler, + LinearMultistepSampler) +from sgm.util import load_model_from_config + + +class ModelArchitecture(str, Enum): + SD_2_1 = "stable-diffusion-v2-1" + SD_2_1_768 = "stable-diffusion-v2-1-768" + SDXL_V0_9_BASE = "stable-diffusion-xl-v0-9-base" + SDXL_V0_9_REFINER = "stable-diffusion-xl-v0-9-refiner" + SDXL_V1_BASE = "stable-diffusion-xl-v1-base" + SDXL_V1_REFINER = "stable-diffusion-xl-v1-refiner" + + +class Sampler(str, Enum): + EULER_EDM = "EulerEDMSampler" + HEUN_EDM = "HeunEDMSampler" + EULER_ANCESTRAL = "EulerAncestralSampler" + DPMPP2S_ANCESTRAL = "DPMPP2SAncestralSampler" + DPMPP2M = "DPMPP2MSampler" + LINEAR_MULTISTEP = "LinearMultistepSampler" + + +class Discretization(str, Enum): + LEGACY_DDPM = "LegacyDDPMDiscretization" + EDM = "EDMDiscretization" + + +class Guider(str, Enum): + VANILLA = "VanillaCFG" + IDENTITY = "IdentityGuider" + + +class Thresholder(str, Enum): + NONE = "None" + + +@dataclass +class SamplingParams: + width: int = 1024 + height: int = 1024 + steps: int = 50 + sampler: Sampler = Sampler.DPMPP2M + discretization: Discretization = Discretization.LEGACY_DDPM + guider: Guider = Guider.VANILLA + thresholder: Thresholder = Thresholder.NONE + scale: float = 6.0 + aesthetic_score: float = 5.0 + negative_aesthetic_score: float = 5.0 + img2img_strength: float = 1.0 + orig_width: int = 1024 + orig_height: int = 1024 + crop_coords_top: int = 0 + crop_coords_left: int = 0 + sigma_min: float = 0.0292 + sigma_max: float = 14.6146 + rho: float = 3.0 + s_churn: float = 0.0 + s_tmin: float = 0.0 + s_tmax: float = 999.0 + s_noise: float = 1.0 + eta: float = 1.0 + order: int = 4 + + +@dataclass +class SamplingSpec: + width: int + height: int + channels: int + factor: int + is_legacy: bool + config: str + ckpt: str + is_guided: bool + + +model_specs = { + ModelArchitecture.SD_2_1: SamplingSpec( + height=512, + width=512, + channels=4, + factor=8, + is_legacy=True, + config="sd_2_1.yaml", + ckpt="v2-1_512-ema-pruned.safetensors", + is_guided=True, + ), + ModelArchitecture.SD_2_1_768: SamplingSpec( + height=768, + width=768, + channels=4, + factor=8, + is_legacy=True, + config="sd_2_1_768.yaml", + ckpt="v2-1_768-ema-pruned.safetensors", + is_guided=True, + ), + ModelArchitecture.SDXL_V0_9_BASE: SamplingSpec( + height=1024, + width=1024, + channels=4, + factor=8, + is_legacy=False, + config="sd_xl_base.yaml", + ckpt="sd_xl_base_0.9.safetensors", + is_guided=True, + ), + ModelArchitecture.SDXL_V0_9_REFINER: SamplingSpec( + height=1024, + width=1024, + channels=4, + factor=8, + is_legacy=True, + config="sd_xl_refiner.yaml", + ckpt="sd_xl_refiner_0.9.safetensors", + is_guided=True, + ), + ModelArchitecture.SDXL_V1_BASE: SamplingSpec( + height=1024, + width=1024, + channels=4, + factor=8, + is_legacy=False, + config="sd_xl_base.yaml", + ckpt="sd_xl_base_1.0.safetensors", + is_guided=True, + ), + ModelArchitecture.SDXL_V1_REFINER: SamplingSpec( + height=1024, + width=1024, + channels=4, + factor=8, + is_legacy=True, + config="sd_xl_refiner.yaml", + ckpt="sd_xl_refiner_1.0.safetensors", + is_guided=True, + ), +} + + +class SamplingPipeline: + def __init__( + self, + model_id: ModelArchitecture, + model_path="checkpoints", + config_path="configs/inference", + device="cuda", + use_fp16=True, + ) -> None: + if model_id not in model_specs: + raise ValueError(f"Model {model_id} not supported") + self.model_id = model_id + self.specs = model_specs[self.model_id] + self.config = str(pathlib.Path(config_path, self.specs.config)) + self.ckpt = str(pathlib.Path(model_path, self.specs.ckpt)) + self.device = device + self.model = self._load_model(device=device, use_fp16=use_fp16) + + def _load_model(self, device="cuda", use_fp16=True): + config = OmegaConf.load(self.config) + model = load_model_from_config(config, self.ckpt) + if model is None: + raise ValueError(f"Model {self.model_id} could not be loaded") + model.to(device) + if use_fp16: + model.conditioner.half() + model.model.half() + return model + + def text_to_image( + self, + params: SamplingParams, + prompt: str, + negative_prompt: str = "", + samples: int = 1, + return_latents: bool = False, + ): + sampler = get_sampler_config(params) + value_dict = asdict(params) + value_dict["prompt"] = prompt + value_dict["negative_prompt"] = negative_prompt + value_dict["target_width"] = params.width + value_dict["target_height"] = params.height + return do_sample( + self.model, + sampler, + value_dict, + samples, + params.height, + params.width, + self.specs.channels, + self.specs.factor, + force_uc_zero_embeddings=["txt"] if not self.specs.is_legacy else [], + return_latents=return_latents, + filter=None, + ) + + def image_to_image( + self, + params: SamplingParams, + image, + prompt: str, + negative_prompt: str = "", + samples: int = 1, + return_latents: bool = False, + ): + sampler = get_sampler_config(params) + + if params.img2img_strength < 1.0: + sampler.discretization = Img2ImgDiscretizationWrapper( + sampler.discretization, + strength=params.img2img_strength, + ) + height, width = image.shape[2], image.shape[3] + value_dict = asdict(params) + value_dict["prompt"] = prompt + value_dict["negative_prompt"] = negative_prompt + value_dict["target_width"] = width + value_dict["target_height"] = height + return do_img2img( + image, + self.model, + sampler, + value_dict, + samples, + force_uc_zero_embeddings=["txt"] if not self.specs.is_legacy else [], + return_latents=return_latents, + filter=None, + ) + + def refiner( + self, + params: SamplingParams, + image, + prompt: str, + negative_prompt: Optional[str] = None, + samples: int = 1, + return_latents: bool = False, + ): + sampler = get_sampler_config(params) + value_dict = { + "orig_width": image.shape[3] * 8, + "orig_height": image.shape[2] * 8, + "target_width": image.shape[3] * 8, + "target_height": image.shape[2] * 8, + "prompt": prompt, + "negative_prompt": negative_prompt, + "crop_coords_top": 0, + "crop_coords_left": 0, + "aesthetic_score": 6.0, + "negative_aesthetic_score": 2.5, + } + + return do_img2img( + image, + self.model, + sampler, + value_dict, + samples, + skip_encode=True, + return_latents=return_latents, + filter=None, + ) + + +def get_guider_config(params: SamplingParams): + if params.guider == Guider.IDENTITY: + guider_config = { + "target": "sgm.modules.diffusionmodules.guiders.IdentityGuider" + } + elif params.guider == Guider.VANILLA: + scale = params.scale + + thresholder = params.thresholder + + if thresholder == Thresholder.NONE: + dyn_thresh_config = { + "target": "sgm.modules.diffusionmodules.sampling_utils.NoDynamicThresholding" + } + else: + raise NotImplementedError + + guider_config = { + "target": "sgm.modules.diffusionmodules.guiders.VanillaCFG", + "params": {"scale": scale, "dyn_thresh_config": dyn_thresh_config}, + } + else: + raise NotImplementedError + return guider_config + + +def get_discretization_config(params: SamplingParams): + if params.discretization == Discretization.LEGACY_DDPM: + discretization_config = { + "target": "sgm.modules.diffusionmodules.discretizer.LegacyDDPMDiscretization", + } + elif params.discretization == Discretization.EDM: + discretization_config = { + "target": "sgm.modules.diffusionmodules.discretizer.EDMDiscretization", + "params": { + "sigma_min": params.sigma_min, + "sigma_max": params.sigma_max, + "rho": params.rho, + }, + } + else: + raise ValueError(f"unknown discretization {params.discretization}") + return discretization_config + + +def get_sampler_config(params: SamplingParams): + discretization_config = get_discretization_config(params) + guider_config = get_guider_config(params) + sampler = None + if params.sampler == Sampler.EULER_EDM: + return EulerEDMSampler( + num_steps=params.steps, + discretization_config=discretization_config, + guider_config=guider_config, + s_churn=params.s_churn, + s_tmin=params.s_tmin, + s_tmax=params.s_tmax, + s_noise=params.s_noise, + verbose=True, + ) + if params.sampler == Sampler.HEUN_EDM: + return HeunEDMSampler( + num_steps=params.steps, + discretization_config=discretization_config, + guider_config=guider_config, + s_churn=params.s_churn, + s_tmin=params.s_tmin, + s_tmax=params.s_tmax, + s_noise=params.s_noise, + verbose=True, + ) + if params.sampler == Sampler.EULER_ANCESTRAL: + return EulerAncestralSampler( + num_steps=params.steps, + discretization_config=discretization_config, + guider_config=guider_config, + eta=params.eta, + s_noise=params.s_noise, + verbose=True, + ) + if params.sampler == Sampler.DPMPP2S_ANCESTRAL: + return DPMPP2SAncestralSampler( + num_steps=params.steps, + discretization_config=discretization_config, + guider_config=guider_config, + eta=params.eta, + s_noise=params.s_noise, + verbose=True, + ) + if params.sampler == Sampler.DPMPP2M: + return DPMPP2MSampler( + num_steps=params.steps, + discretization_config=discretization_config, + guider_config=guider_config, + verbose=True, + ) + if params.sampler == Sampler.LINEAR_MULTISTEP: + return LinearMultistepSampler( + num_steps=params.steps, + discretization_config=discretization_config, + guider_config=guider_config, + order=params.order, + verbose=True, + ) + + raise ValueError(f"unknown sampler {params.sampler}!") diff --git a/sgm/inference/helpers.py b/sgm/inference/helpers.py new file mode 100644 index 0000000..31b0ec3 --- /dev/null +++ b/sgm/inference/helpers.py @@ -0,0 +1,305 @@ +import math +import os +from typing import List, Optional, Union + +import numpy as np +import torch +from einops import rearrange +from imwatermark import WatermarkEncoder +from omegaconf import ListConfig +from PIL import Image +from torch import autocast + +from sgm.util import append_dims + + +class WatermarkEmbedder: + def __init__(self, watermark): + self.watermark = watermark + self.num_bits = len(WATERMARK_BITS) + self.encoder = WatermarkEncoder() + self.encoder.set_watermark("bits", self.watermark) + + def __call__(self, image: torch.Tensor) -> torch.Tensor: + """ + Adds a predefined watermark to the input image + + Args: + image: ([N,] B, RGB, H, W) in range [0, 1] + + Returns: + same as input but watermarked + """ + squeeze = len(image.shape) == 4 + if squeeze: + image = image[None, ...] + n = image.shape[0] + image_np = rearrange( + (255 * image).detach().cpu(), "n b c h w -> (n b) h w c" + ).numpy()[:, :, :, ::-1] + # torch (b, c, h, w) in [0, 1] -> numpy (b, h, w, c) [0, 255] + # watermarking libary expects input as cv2 BGR format + for k in range(image_np.shape[0]): + image_np[k] = self.encoder.encode(image_np[k], "dwtDct") + image = torch.from_numpy( + rearrange(image_np[:, :, :, ::-1], "(n b) h w c -> n b c h w", n=n) + ).to(image.device) + image = torch.clamp(image / 255, min=0.0, max=1.0) + if squeeze: + image = image[0] + return image + + +# A fixed 48-bit message that was choosen at random +# WATERMARK_MESSAGE = 0xB3EC907BB19E +WATERMARK_MESSAGE = 0b101100111110110010010000011110111011000110011110 +# bin(x)[2:] gives bits of x as str, use int to convert them to 0/1 +WATERMARK_BITS = [int(bit) for bit in bin(WATERMARK_MESSAGE)[2:]] +embed_watermark = WatermarkEmbedder(WATERMARK_BITS) + + +def get_unique_embedder_keys_from_conditioner(conditioner): + return list({x.input_key for x in conditioner.embedders}) + + +def perform_save_locally(save_path, samples): + os.makedirs(os.path.join(save_path), exist_ok=True) + base_count = len(os.listdir(os.path.join(save_path))) + samples = embed_watermark(samples) + for sample in samples: + sample = 255.0 * rearrange(sample.cpu().numpy(), "c h w -> h w c") + Image.fromarray(sample.astype(np.uint8)).save( + os.path.join(save_path, f"{base_count:09}.png") + ) + base_count += 1 + + +class Img2ImgDiscretizationWrapper: + """ + wraps a discretizer, and prunes the sigmas + params: + strength: float between 0.0 and 1.0. 1.0 means full sampling (all sigmas are returned) + """ + + def __init__(self, discretization, strength: float = 1.0): + self.discretization = discretization + self.strength = strength + assert 0.0 <= self.strength <= 1.0 + + def __call__(self, *args, **kwargs): + # sigmas start large first, and decrease then + sigmas = self.discretization(*args, **kwargs) + print(f"sigmas after discretization, before pruning img2img: ", sigmas) + sigmas = torch.flip(sigmas, (0,)) + sigmas = sigmas[: max(int(self.strength * len(sigmas)), 1)] + print("prune index:", max(int(self.strength * len(sigmas)), 1)) + sigmas = torch.flip(sigmas, (0,)) + print(f"sigmas after pruning: ", sigmas) + return sigmas + + +def do_sample( + model, + sampler, + value_dict, + num_samples, + H, + W, + C, + F, + force_uc_zero_embeddings: Optional[List] = None, + batch2model_input: Optional[List] = None, + return_latents=False, + filter=None, + device="cuda", +): + if force_uc_zero_embeddings is None: + force_uc_zero_embeddings = [] + if batch2model_input is None: + batch2model_input = [] + + with torch.no_grad(): + with autocast(device) as precision_scope: + with model.ema_scope(): + num_samples = [num_samples] + batch, batch_uc = get_batch( + get_unique_embedder_keys_from_conditioner(model.conditioner), + value_dict, + num_samples, + ) + for key in batch: + if isinstance(batch[key], torch.Tensor): + print(key, batch[key].shape) + elif isinstance(batch[key], list): + print(key, [len(l) for l in batch[key]]) + else: + print(key, batch[key]) + c, uc = model.conditioner.get_unconditional_conditioning( + batch, + batch_uc=batch_uc, + force_uc_zero_embeddings=force_uc_zero_embeddings, + ) + + for k in c: + if not k == "crossattn": + c[k], uc[k] = map( + lambda y: y[k][: math.prod(num_samples)].to(device), (c, uc) + ) + + additional_model_inputs = {} + for k in batch2model_input: + additional_model_inputs[k] = batch[k] + + shape = (math.prod(num_samples), C, H // F, W // F) + randn = torch.randn(shape).to(device) + + def denoiser(input, sigma, c): + return model.denoiser( + model.model, input, sigma, c, **additional_model_inputs + ) + + samples_z = sampler(denoiser, randn, cond=c, uc=uc) + samples_x = model.decode_first_stage(samples_z) + samples = torch.clamp((samples_x + 1.0) / 2.0, min=0.0, max=1.0) + + if filter is not None: + samples = filter(samples) + + if return_latents: + return samples, samples_z + return samples + + +def get_batch(keys, value_dict, N: Union[List, ListConfig], device="cuda"): + # Hardcoded demo setups; might undergo some changes in the future + + batch = {} + batch_uc = {} + + for key in keys: + if key == "txt": + batch["txt"] = ( + np.repeat([value_dict["prompt"]], repeats=math.prod(N)) + .reshape(N) + .tolist() + ) + batch_uc["txt"] = ( + np.repeat([value_dict["negative_prompt"]], repeats=math.prod(N)) + .reshape(N) + .tolist() + ) + elif key == "original_size_as_tuple": + batch["original_size_as_tuple"] = ( + torch.tensor([value_dict["orig_height"], value_dict["orig_width"]]) + .to(device) + .repeat(*N, 1) + ) + elif key == "crop_coords_top_left": + batch["crop_coords_top_left"] = ( + torch.tensor( + [value_dict["crop_coords_top"], value_dict["crop_coords_left"]] + ) + .to(device) + .repeat(*N, 1) + ) + elif key == "aesthetic_score": + batch["aesthetic_score"] = ( + torch.tensor([value_dict["aesthetic_score"]]).to(device).repeat(*N, 1) + ) + batch_uc["aesthetic_score"] = ( + torch.tensor([value_dict["negative_aesthetic_score"]]) + .to(device) + .repeat(*N, 1) + ) + + elif key == "target_size_as_tuple": + batch["target_size_as_tuple"] = ( + torch.tensor([value_dict["target_height"], value_dict["target_width"]]) + .to(device) + .repeat(*N, 1) + ) + else: + batch[key] = value_dict[key] + + for key in batch.keys(): + if key not in batch_uc and isinstance(batch[key], torch.Tensor): + batch_uc[key] = torch.clone(batch[key]) + return batch, batch_uc + + +def get_input_image_tensor(image: Image.Image, device="cuda"): + w, h = image.size + print(f"loaded input image of size ({w}, {h})") + width, height = map( + lambda x: x - x % 64, (w, h) + ) # resize to integer multiple of 64 + image = image.resize((width, height)) + image_array = np.array(image.convert("RGB")) + image_array = image_array[None].transpose(0, 3, 1, 2) + image_tensor = torch.from_numpy(image_array).to(dtype=torch.float32) / 127.5 - 1.0 + return image_tensor.to(device) + + +def do_img2img( + img, + model, + sampler, + value_dict, + num_samples, + force_uc_zero_embeddings=[], + additional_kwargs={}, + offset_noise_level: float = 0.0, + return_latents=False, + skip_encode=False, + filter=None, + device="cuda", +): + with torch.no_grad(): + with autocast(device) as precision_scope: + with model.ema_scope(): + batch, batch_uc = get_batch( + get_unique_embedder_keys_from_conditioner(model.conditioner), + value_dict, + [num_samples], + ) + c, uc = model.conditioner.get_unconditional_conditioning( + batch, + batch_uc=batch_uc, + force_uc_zero_embeddings=force_uc_zero_embeddings, + ) + + for k in c: + c[k], uc[k] = map(lambda y: y[k][:num_samples].to(device), (c, uc)) + + for k in additional_kwargs: + c[k] = uc[k] = additional_kwargs[k] + if skip_encode: + z = img + else: + z = model.encode_first_stage(img) + noise = torch.randn_like(z) + sigmas = sampler.discretization(sampler.num_steps) + sigma = sigmas[0].to(z.device) + + if offset_noise_level > 0.0: + noise = noise + offset_noise_level * append_dims( + torch.randn(z.shape[0], device=z.device), z.ndim + ) + noised_z = z + noise * append_dims(sigma, z.ndim) + noised_z = noised_z / torch.sqrt( + 1.0 + sigmas[0] ** 2.0 + ) # Note: hardcoded to DDPM-like scaling. need to generalize later. + + def denoiser(x, sigma, c): + return model.denoiser(model.model, x, sigma, c) + + samples_z = sampler(denoiser, noised_z, cond=c, uc=uc) + samples_x = model.decode_first_stage(samples_z) + samples = torch.clamp((samples_x + 1.0) / 2.0, min=0.0, max=1.0) + + if filter is not None: + samples = filter(samples) + + if return_latents: + return samples, samples_z + return samples diff --git a/sgm/lr_scheduler.py b/sgm/lr_scheduler.py new file mode 100644 index 0000000..b2f4d38 --- /dev/null +++ b/sgm/lr_scheduler.py @@ -0,0 +1,135 @@ +import numpy as np + + +class LambdaWarmUpCosineScheduler: + """ + note: use with a base_lr of 1.0 + """ + + def __init__( + self, + warm_up_steps, + lr_min, + lr_max, + lr_start, + max_decay_steps, + verbosity_interval=0, + ): + self.lr_warm_up_steps = warm_up_steps + self.lr_start = lr_start + self.lr_min = lr_min + self.lr_max = lr_max + self.lr_max_decay_steps = max_decay_steps + self.last_lr = 0.0 + self.verbosity_interval = verbosity_interval + + def schedule(self, n, **kwargs): + if self.verbosity_interval > 0: + if n % self.verbosity_interval == 0: + print(f"current step: {n}, recent lr-multiplier: {self.last_lr}") + if n < self.lr_warm_up_steps: + lr = ( + self.lr_max - self.lr_start + ) / self.lr_warm_up_steps * n + self.lr_start + self.last_lr = lr + return lr + else: + t = (n - self.lr_warm_up_steps) / ( + self.lr_max_decay_steps - self.lr_warm_up_steps + ) + t = min(t, 1.0) + lr = self.lr_min + 0.5 * (self.lr_max - self.lr_min) * ( + 1 + np.cos(t * np.pi) + ) + self.last_lr = lr + return lr + + def __call__(self, n, **kwargs): + return self.schedule(n, **kwargs) + + +class LambdaWarmUpCosineScheduler2: + """ + supports repeated iterations, configurable via lists + note: use with a base_lr of 1.0. + """ + + def __init__( + self, warm_up_steps, f_min, f_max, f_start, cycle_lengths, verbosity_interval=0 + ): + assert ( + len(warm_up_steps) + == len(f_min) + == len(f_max) + == len(f_start) + == len(cycle_lengths) + ) + self.lr_warm_up_steps = warm_up_steps + self.f_start = f_start + self.f_min = f_min + self.f_max = f_max + self.cycle_lengths = cycle_lengths + self.cum_cycles = np.cumsum([0] + list(self.cycle_lengths)) + self.last_f = 0.0 + self.verbosity_interval = verbosity_interval + + def find_in_interval(self, n): + interval = 0 + for cl in self.cum_cycles[1:]: + if n <= cl: + return interval + interval += 1 + + def schedule(self, n, **kwargs): + cycle = self.find_in_interval(n) + n = n - self.cum_cycles[cycle] + if self.verbosity_interval > 0: + if n % self.verbosity_interval == 0: + print( + f"current step: {n}, recent lr-multiplier: {self.last_f}, " + f"current cycle {cycle}" + ) + if n < self.lr_warm_up_steps[cycle]: + f = (self.f_max[cycle] - self.f_start[cycle]) / self.lr_warm_up_steps[ + cycle + ] * n + self.f_start[cycle] + self.last_f = f + return f + else: + t = (n - self.lr_warm_up_steps[cycle]) / ( + self.cycle_lengths[cycle] - self.lr_warm_up_steps[cycle] + ) + t = min(t, 1.0) + f = self.f_min[cycle] + 0.5 * (self.f_max[cycle] - self.f_min[cycle]) * ( + 1 + np.cos(t * np.pi) + ) + self.last_f = f + return f + + def __call__(self, n, **kwargs): + return self.schedule(n, **kwargs) + + +class LambdaLinearScheduler(LambdaWarmUpCosineScheduler2): + def schedule(self, n, **kwargs): + cycle = self.find_in_interval(n) + n = n - self.cum_cycles[cycle] + if self.verbosity_interval > 0: + if n % self.verbosity_interval == 0: + print( + f"current step: {n}, recent lr-multiplier: {self.last_f}, " + f"current cycle {cycle}" + ) + + if n < self.lr_warm_up_steps[cycle]: + f = (self.f_max[cycle] - self.f_start[cycle]) / self.lr_warm_up_steps[ + cycle + ] * n + self.f_start[cycle] + self.last_f = f + return f + else: + f = self.f_min[cycle] + (self.f_max[cycle] - self.f_min[cycle]) * ( + self.cycle_lengths[cycle] - n + ) / (self.cycle_lengths[cycle]) + self.last_f = f + return f diff --git a/sgm/models/__init__.py b/sgm/models/__init__.py new file mode 100644 index 0000000..c410b37 --- /dev/null +++ b/sgm/models/__init__.py @@ -0,0 +1,2 @@ +from .autoencoder import AutoencodingEngine +from .diffusion import DiffusionEngine diff --git a/sgm/models/autoencoder.py b/sgm/models/autoencoder.py new file mode 100644 index 0000000..ebf56a6 --- /dev/null +++ b/sgm/models/autoencoder.py @@ -0,0 +1,679 @@ +import logging +import math +import re +from abc import abstractmethod +from contextlib import contextmanager +from typing import Any, Dict, List, Optional, Tuple, Union + +import pytorch_lightning as pl +import torch +import torch.nn as nn +from einops import rearrange +from packaging import version +from safetensors.torch import load_file + +from ..modules.autoencoding.regularizers import AbstractRegularizer +from ..modules.ema import LitEma +from ..util import (default, get_nested_attribute, get_obj_from_str, + instantiate_from_config) + +logpy = logging.getLogger(__name__) + + +class AbstractAutoencoder(pl.LightningModule): + """ + This is the base class for all autoencoders, including image autoencoders, image autoencoders with discriminators, + unCLIP models, etc. Hence, it is fairly general, and specific features + (e.g. discriminator training, encoding, decoding) must be implemented in subclasses. + """ + + def __init__( + self, + ema_decay: Union[None, float] = None, + monitor: Union[None, str] = None, + input_key: str = "jpg", + ): + super().__init__() + + self.input_key = input_key + self.use_ema = ema_decay is not None + if monitor is not None: + self.monitor = monitor + + if self.use_ema: + self.model_ema = LitEma(self, decay=ema_decay) + logpy.info(f"Keeping EMAs of {len(list(self.model_ema.buffers()))}.") + + if version.parse(torch.__version__) >= version.parse("2.0.0"): + self.automatic_optimization = False + + def apply_ckpt(self, ckpt: Union[None, str, dict]): + if ckpt is None: + return + if ckpt.endswith("ckpt"): + sd = torch.load(ckpt, map_location="cpu")["state_dict"] + elif ckpt.endswith("safetensors"): + sd = load_file(ckpt) + else: + raise NotImplementedError + + missing, unexpected = self.load_state_dict(sd, strict=False) + print( + f"Restored from {ckpt} with {len(missing)} missing and {len(unexpected)} unexpected keys" + ) + + @abstractmethod + def get_input(self, batch) -> Any: + raise NotImplementedError() + + def on_train_batch_end(self, *args, **kwargs): + # for EMA computation + if self.use_ema: + self.model_ema(self) + + @contextmanager + def ema_scope(self, context=None): + if self.use_ema: + self.model_ema.store(self.parameters()) + self.model_ema.copy_to(self) + if context is not None: + logpy.info(f"{context}: Switched to EMA weights") + try: + yield None + finally: + if self.use_ema: + self.model_ema.restore(self.parameters()) + if context is not None: + logpy.info(f"{context}: Restored training weights") + + @abstractmethod + def encode(self, *args, **kwargs) -> torch.Tensor: + raise NotImplementedError("encode()-method of abstract base class called") + + @abstractmethod + def decode(self, *args, **kwargs) -> torch.Tensor: + raise NotImplementedError("decode()-method of abstract base class called") + + def instantiate_optimizer_from_config(self, params, lr, cfg): + logpy.info(f"loading >>> {cfg['target']} <<< optimizer from config") + return get_obj_from_str(cfg["target"])( + params, lr=lr, **cfg.get("params", dict()) + ) + + def configure_optimizers(self) -> Any: + raise NotImplementedError() + + +class AutoencodingEngine(AbstractAutoencoder): + """ + Base class for all image autoencoders that we train, like VQGAN or AutoencoderKL + (we also restore them explicitly as special cases for legacy reasons). + Regularizations such as KL or VQ are moved to the regularizer class. + """ + + def __init__( + self, + *args, + encoder_config: Dict, + decoder_config: Dict, + loss_config: Dict, + regularizer_config: Dict, + optimizer_config: Union[Dict, None] = None, + lr_g_factor: float = 1.0, + trainable_ae_params: Optional[List[List[str]]] = None, + ae_optimizer_args: Optional[List[dict]] = None, + trainable_disc_params: Optional[List[List[str]]] = None, + disc_optimizer_args: Optional[List[dict]] = None, + disc_start_iter: int = 0, + diff_boost_factor: float = 3.0, + ckpt_engine: Union[None, str, dict] = None, + ckpt_path: Optional[str] = None, + additional_decode_keys: Optional[List[str]] = None, + **kwargs, + ): + super().__init__(*args, **kwargs) + self.automatic_optimization = False # pytorch lightning + + self.encoder: torch.nn.Module = instantiate_from_config(encoder_config) + self.decoder: torch.nn.Module = instantiate_from_config(decoder_config) + self.loss: torch.nn.Module = instantiate_from_config(loss_config) + self.regularization: AbstractRegularizer = instantiate_from_config( + regularizer_config + ) + self.optimizer_config = default( + optimizer_config, {"target": "torch.optim.Adam"} + ) + self.diff_boost_factor = diff_boost_factor + self.disc_start_iter = disc_start_iter + self.lr_g_factor = lr_g_factor + self.trainable_ae_params = trainable_ae_params + if self.trainable_ae_params is not None: + self.ae_optimizer_args = default( + ae_optimizer_args, + [{} for _ in range(len(self.trainable_ae_params))], + ) + assert len(self.ae_optimizer_args) == len(self.trainable_ae_params) + else: + self.ae_optimizer_args = [{}] # makes type consitent + + self.trainable_disc_params = trainable_disc_params + if self.trainable_disc_params is not None: + self.disc_optimizer_args = default( + disc_optimizer_args, + [{} for _ in range(len(self.trainable_disc_params))], + ) + assert len(self.disc_optimizer_args) == len(self.trainable_disc_params) + else: + self.disc_optimizer_args = [{}] # makes type consitent + + if ckpt_path is not None: + assert ckpt_engine is None, "Can't set ckpt_engine and ckpt_path" + logpy.warn("Checkpoint path is deprecated, use `checkpoint_egnine` instead") + self.apply_ckpt(default(ckpt_path, ckpt_engine)) + self.additional_decode_keys = set(default(additional_decode_keys, [])) + + def get_input(self, batch: Dict) -> torch.Tensor: + # assuming unified data format, dataloader returns a dict. + # image tensors should be scaled to -1 ... 1 and in channels-first + # format (e.g., bchw instead if bhwc) + return batch[self.input_key] + + def get_autoencoder_params(self) -> list: + params = [] + if hasattr(self.loss, "get_trainable_autoencoder_parameters"): + params += list(self.loss.get_trainable_autoencoder_parameters()) + if hasattr(self.regularization, "get_trainable_parameters"): + params += list(self.regularization.get_trainable_parameters()) + params = params + list(self.encoder.parameters()) + params = params + list(self.decoder.parameters()) + return params + + def get_discriminator_params(self) -> list: + if hasattr(self.loss, "get_trainable_parameters"): + params = list(self.loss.get_trainable_parameters()) # e.g., discriminator + else: + params = [] + return params + + def get_last_layer(self): + return self.decoder.get_last_layer() + + def encode( + self, + x: torch.Tensor, + return_reg_log: bool = False, + unregularized: bool = False, + ) -> Union[torch.Tensor, Tuple[torch.Tensor, dict]]: + z = self.encoder(x) + if unregularized: + return z, dict() + z, reg_log = self.regularization(z) + if return_reg_log: + return z, reg_log + return z + + def decode(self, z: torch.Tensor, **kwargs) -> torch.Tensor: + x = self.decoder(z, **kwargs) + return x + + def forward( + self, x: torch.Tensor, **additional_decode_kwargs + ) -> Tuple[torch.Tensor, torch.Tensor, dict]: + z, reg_log = self.encode(x, return_reg_log=True) + dec = self.decode(z, **additional_decode_kwargs) + return z, dec, reg_log + + def inner_training_step( + self, batch: dict, batch_idx: int, optimizer_idx: int = 0 + ) -> torch.Tensor: + x = self.get_input(batch) + additional_decode_kwargs = { + key: batch[key] for key in self.additional_decode_keys.intersection(batch) + } + z, xrec, regularization_log = self(x, **additional_decode_kwargs) + if hasattr(self.loss, "forward_keys"): + extra_info = { + "z": z, + "optimizer_idx": optimizer_idx, + "global_step": self.global_step, + "last_layer": self.get_last_layer(), + "split": "train", + "regularization_log": regularization_log, + "autoencoder": self, + } + extra_info = {k: extra_info[k] for k in self.loss.forward_keys} + else: + extra_info = dict() + + if optimizer_idx == 0: + # autoencode + out_loss = self.loss(x, xrec, **extra_info) + if isinstance(out_loss, tuple): + aeloss, log_dict_ae = out_loss + else: + # simple loss function + aeloss = out_loss + log_dict_ae = {"train/loss/rec": aeloss.detach()} + + self.log_dict( + log_dict_ae, + prog_bar=False, + logger=True, + on_step=True, + on_epoch=True, + sync_dist=False, + ) + self.log( + "loss", + aeloss.mean().detach(), + prog_bar=True, + logger=False, + on_epoch=False, + on_step=True, + ) + return aeloss + elif optimizer_idx == 1: + # discriminator + discloss, log_dict_disc = self.loss(x, xrec, **extra_info) + # -> discriminator always needs to return a tuple + self.log_dict( + log_dict_disc, prog_bar=False, logger=True, on_step=True, on_epoch=True + ) + return discloss + else: + raise NotImplementedError(f"Unknown optimizer {optimizer_idx}") + + def training_step(self, batch: dict, batch_idx: int): + opts = self.optimizers() + if not isinstance(opts, list): + # Non-adversarial case + opts = [opts] + optimizer_idx = batch_idx % len(opts) + if self.global_step < self.disc_start_iter: + optimizer_idx = 0 + opt = opts[optimizer_idx] + opt.zero_grad() + with opt.toggle_model(): + loss = self.inner_training_step( + batch, batch_idx, optimizer_idx=optimizer_idx + ) + self.manual_backward(loss) + opt.step() + + def validation_step(self, batch: dict, batch_idx: int) -> Dict: + log_dict = self._validation_step(batch, batch_idx) + with self.ema_scope(): + log_dict_ema = self._validation_step(batch, batch_idx, postfix="_ema") + log_dict.update(log_dict_ema) + return log_dict + + def _validation_step(self, batch: dict, batch_idx: int, postfix: str = "") -> Dict: + x = self.get_input(batch) + + z, xrec, regularization_log = self(x) + if hasattr(self.loss, "forward_keys"): + extra_info = { + "z": z, + "optimizer_idx": 0, + "global_step": self.global_step, + "last_layer": self.get_last_layer(), + "split": "val" + postfix, + "regularization_log": regularization_log, + "autoencoder": self, + } + extra_info = {k: extra_info[k] for k in self.loss.forward_keys} + else: + extra_info = dict() + out_loss = self.loss(x, xrec, **extra_info) + if isinstance(out_loss, tuple): + aeloss, log_dict_ae = out_loss + else: + # simple loss function + aeloss = out_loss + log_dict_ae = {f"val{postfix}/loss/rec": aeloss.detach()} + full_log_dict = log_dict_ae + + if "optimizer_idx" in extra_info: + extra_info["optimizer_idx"] = 1 + discloss, log_dict_disc = self.loss(x, xrec, **extra_info) + full_log_dict.update(log_dict_disc) + self.log( + f"val{postfix}/loss/rec", + log_dict_ae[f"val{postfix}/loss/rec"], + sync_dist=True, + ) + self.log_dict(full_log_dict, sync_dist=True) + return full_log_dict + + def get_param_groups( + self, parameter_names: List[List[str]], optimizer_args: List[dict] + ) -> Tuple[List[Dict[str, Any]], int]: + groups = [] + num_params = 0 + for names, args in zip(parameter_names, optimizer_args): + params = [] + for pattern_ in names: + pattern_params = [] + pattern = re.compile(pattern_) + for p_name, param in self.named_parameters(): + if re.match(pattern, p_name): + pattern_params.append(param) + num_params += param.numel() + if len(pattern_params) == 0: + logpy.warn(f"Did not find parameters for pattern {pattern_}") + params.extend(pattern_params) + groups.append({"params": params, **args}) + return groups, num_params + + def configure_optimizers(self) -> List[torch.optim.Optimizer]: + if self.trainable_ae_params is None: + ae_params = self.get_autoencoder_params() + else: + ae_params, num_ae_params = self.get_param_groups( + self.trainable_ae_params, self.ae_optimizer_args + ) + logpy.info(f"Number of trainable autoencoder parameters: {num_ae_params:,}") + if self.trainable_disc_params is None: + disc_params = self.get_discriminator_params() + else: + disc_params, num_disc_params = self.get_param_groups( + self.trainable_disc_params, self.disc_optimizer_args + ) + logpy.info( + f"Number of trainable discriminator parameters: {num_disc_params:,}" + ) + opt_ae = self.instantiate_optimizer_from_config( + ae_params, + default(self.lr_g_factor, 1.0) * self.learning_rate, + self.optimizer_config, + ) + opts = [opt_ae] + if len(disc_params) > 0: + opt_disc = self.instantiate_optimizer_from_config( + disc_params, self.learning_rate, self.optimizer_config + ) + opts.append(opt_disc) + + return opts + + @torch.no_grad() + def log_images( + self, batch: dict, additional_log_kwargs: Optional[Dict] = None, **kwargs + ) -> dict: + log = dict() + additional_decode_kwargs = {} + x = self.get_input(batch) + additional_decode_kwargs.update( + {key: batch[key] for key in self.additional_decode_keys.intersection(batch)} + ) + + _, xrec, _ = self(x, **additional_decode_kwargs) + log["inputs"] = x + log["reconstructions"] = xrec + diff = 0.5 * torch.abs(torch.clamp(xrec, -1.0, 1.0) - x) + diff.clamp_(0, 1.0) + log["diff"] = 2.0 * diff - 1.0 + # diff_boost shows location of small errors, by boosting their + # brightness. + log["diff_boost"] = ( + 2.0 * torch.clamp(self.diff_boost_factor * diff, 0.0, 1.0) - 1 + ) + if hasattr(self.loss, "log_images"): + log.update(self.loss.log_images(x, xrec)) + with self.ema_scope(): + _, xrec_ema, _ = self(x, **additional_decode_kwargs) + log["reconstructions_ema"] = xrec_ema + diff_ema = 0.5 * torch.abs(torch.clamp(xrec_ema, -1.0, 1.0) - x) + diff_ema.clamp_(0, 1.0) + log["diff_ema"] = 2.0 * diff_ema - 1.0 + log["diff_boost_ema"] = ( + 2.0 * torch.clamp(self.diff_boost_factor * diff_ema, 0.0, 1.0) - 1 + ) + if additional_log_kwargs: + additional_decode_kwargs.update(additional_log_kwargs) + _, xrec_add, _ = self(x, **additional_decode_kwargs) + log_str = "reconstructions-" + "-".join( + [f"{key}={additional_log_kwargs[key]}" for key in additional_log_kwargs] + ) + log[log_str] = xrec_add + return log + + +class AutoencodingEngineLegacy(AutoencodingEngine): + def __init__(self, embed_dim: int, **kwargs): + self.max_batch_size = kwargs.pop("max_batch_size", None) + ddconfig = kwargs.pop("ddconfig") + ckpt_path = kwargs.pop("ckpt_path", None) + ckpt_engine = kwargs.pop("ckpt_engine", None) + super().__init__( + encoder_config={ + "target": "sgm.modules.diffusionmodules.model.Encoder", + "params": ddconfig, + }, + decoder_config={ + "target": "sgm.modules.diffusionmodules.model.Decoder", + "params": ddconfig, + }, + **kwargs, + ) + self.quant_conv = torch.nn.Conv2d( + (1 + ddconfig["double_z"]) * ddconfig["z_channels"], + (1 + ddconfig["double_z"]) * embed_dim, + 1, + ) + self.post_quant_conv = torch.nn.Conv2d(embed_dim, ddconfig["z_channels"], 1) + self.embed_dim = embed_dim + + self.apply_ckpt(default(ckpt_path, ckpt_engine)) + + def get_autoencoder_params(self) -> list: + params = super().get_autoencoder_params() + return params + + def encode( + self, x: torch.Tensor, return_reg_log: bool = False + ) -> Union[torch.Tensor, Tuple[torch.Tensor, dict]]: + if self.max_batch_size is None: + z = self.encoder(x) + z = self.quant_conv(z) + else: + N = x.shape[0] + bs = self.max_batch_size + n_batches = int(math.ceil(N / bs)) + z = list() + for i_batch in range(n_batches): + z_batch = self.encoder(x[i_batch * bs : (i_batch + 1) * bs]) + z_batch = self.quant_conv(z_batch) + z.append(z_batch) + z = torch.cat(z, 0) + + z, reg_log = self.regularization(z) + if return_reg_log: + return z, reg_log + return z + + def decode(self, z: torch.Tensor, **decoder_kwargs) -> torch.Tensor: + if self.max_batch_size is None: + dec = self.post_quant_conv(z) + dec = self.decoder(dec, **decoder_kwargs) + else: + N = z.shape[0] + bs = self.max_batch_size + n_batches = int(math.ceil(N / bs)) + dec = list() + for i_batch in range(n_batches): + dec_batch = self.post_quant_conv(z[i_batch * bs : (i_batch + 1) * bs]) + dec_batch = self.decoder(dec_batch, **decoder_kwargs) + dec.append(dec_batch) + dec = torch.cat(dec, 0) + + return dec + + +class AutoencoderKL(AutoencodingEngineLegacy): + def __init__(self, **kwargs): + if "lossconfig" in kwargs: + kwargs["loss_config"] = kwargs.pop("lossconfig") + super().__init__( + regularizer_config={ + "target": ( + "sgm.modules.autoencoding.regularizers" + ".DiagonalGaussianRegularizer" + ) + }, + **kwargs, + ) + +class SevaAutoencoderKL(AutoencodingEngineLegacy): + def __init__(self, **kwargs): + if "lossconfig" in kwargs: + kwargs["loss_config"] = kwargs.pop("lossconfig") + super().__init__( + regularizer_config={ + "target": ( + "sgm.modules.autoencoding.regularizers" + ".DiagonalGaussianRegularizer" + ) + }, + **kwargs, + ) + def encode( + self, x: torch.Tensor, return_reg_log: bool = False + ) -> Union[torch.Tensor, Tuple[torch.Tensor, dict]]: + batch_size, num_frames, num_channels, height, width = x.shape + x = rearrange(x, "b f c h w -> (b f) c h w") + if self.max_batch_size is None: + z = self.encoder(x) + z = self.quant_conv(z) + else: + N = x.shape[0] + bs = self.max_batch_size + n_batches = int(math.ceil(N / bs)) + z = list() + for i_batch in range(n_batches): + z_batch = self.encoder(x[i_batch * bs : (i_batch + 1) * bs]) + z_batch = self.quant_conv(z_batch) + z.append(z_batch) + z = torch.cat(z, 0) + + z, reg_log = self.regularization(z) + z = rearrange(z, "(b f) c h w -> b f c h w", b=batch_size, f=num_frames) + if return_reg_log: + return z, reg_log + return z + + def decode(self, z: torch.Tensor, **decoder_kwargs) -> torch.Tensor: + batch_size, num_frames, num_channels, height, width = z.shape + z = rearrange(z, "b f c h w -> (b f) c h w", b=batch_size, f=num_frames) + + if self.max_batch_size is None: + dec = self.post_quant_conv(z) + dec = self.decoder(dec, **decoder_kwargs) + else: + N = z.shape[0] + bs = self.max_batch_size + n_batches = int(math.ceil(N / bs)) + dec = list() + for i_batch in range(n_batches): + dec_batch = self.post_quant_conv(z[i_batch * bs : (i_batch + 1) * bs]) + dec_batch = self.decoder(dec_batch, **decoder_kwargs) + dec.append(dec_batch) + dec = torch.cat(dec, 0) + dec = rearrange(dec, "(b s) c h w -> b s c h w", b=batch_size, s=num_frames) + return dec + +class AutoencoderLegacyVQ(AutoencodingEngineLegacy): + def __init__( + self, + embed_dim: int, + n_embed: int, + sane_index_shape: bool = False, + **kwargs, + ): + if "lossconfig" in kwargs: + logpy.warn(f"Parameter `lossconfig` is deprecated, use `loss_config`.") + kwargs["loss_config"] = kwargs.pop("lossconfig") + super().__init__( + regularizer_config={ + "target": ( + "sgm.modules.autoencoding.regularizers.quantize" ".VectorQuantizer" + ), + "params": { + "n_e": n_embed, + "e_dim": embed_dim, + "sane_index_shape": sane_index_shape, + }, + }, + **kwargs, + ) + + +class IdentityFirstStage(AbstractAutoencoder): + def __init__(self, *args, **kwargs): + super().__init__(*args, **kwargs) + self.encoder = lambda x: x + self.decoder = lambda x: x + + def get_input(self, x: Any) -> Any: + return x + + def encode(self, x: Any, *args, **kwargs) -> Any: + return self.encoder(x) + + def decode(self, x: Any, *args, **kwargs) -> Any: + return self.decoder(x) + + +class AEIntegerWrapper(nn.Module): + def __init__( + self, + model: nn.Module, + shape: Union[None, Tuple[int, int], List[int]] = (16, 16), + regularization_key: str = "regularization", + encoder_kwargs: Optional[Dict[str, Any]] = None, + ): + super().__init__() + self.model = model + assert hasattr(model, "encode") and hasattr( + model, "decode" + ), "Need AE interface" + self.regularization = get_nested_attribute(model, regularization_key) + self.shape = shape + self.encoder_kwargs = default(encoder_kwargs, {"return_reg_log": True}) + + def encode(self, x) -> torch.Tensor: + assert ( + not self.training + ), f"{self.__class__.__name__} only supports inference currently" + _, log = self.model.encode(x, **self.encoder_kwargs) + assert isinstance(log, dict) + inds = log["min_encoding_indices"] + return rearrange(inds, "b ... -> b (...)") + + def decode( + self, inds: torch.Tensor, shape: Union[None, tuple, list] = None + ) -> torch.Tensor: + # expect inds shape (b, s) with s = h*w + shape = default(shape, self.shape) # Optional[(h, w)] + if shape is not None: + assert len(shape) == 2, f"Unhandeled shape {shape}" + inds = rearrange(inds, "b (h w) -> b h w", h=shape[0], w=shape[1]) + h = self.regularization.get_codebook_entry(inds) # (b, h, w, c) + h = rearrange(h, "b h w c -> b c h w") + return self.model.decode(h) + + +class AutoencoderKLModeOnly(AutoencodingEngineLegacy): + def __init__(self, **kwargs): + if "lossconfig" in kwargs: + kwargs["loss_config"] = kwargs.pop("lossconfig") + super().__init__( + regularizer_config={ + "target": ( + "sgm.modules.autoencoding.regularizers" + ".DiagonalGaussianRegularizer" + ), + "params": {"sample": False}, + }, + **kwargs, + ) diff --git a/sgm/models/diffusion.py b/sgm/models/diffusion.py new file mode 100644 index 0000000..2f3efd3 --- /dev/null +++ b/sgm/models/diffusion.py @@ -0,0 +1,341 @@ +import math +from contextlib import contextmanager +from typing import Any, Dict, List, Optional, Tuple, Union + +import pytorch_lightning as pl +import torch +from omegaconf import ListConfig, OmegaConf +from safetensors.torch import load_file as load_safetensors +from torch.optim.lr_scheduler import LambdaLR + +from ..modules import UNCONDITIONAL_CONFIG +from ..modules.autoencoding.temporal_ae import VideoDecoder +from ..modules.diffusionmodules.wrappers import OPENAIUNETWRAPPER +from ..modules.ema import LitEma +from ..util import (default, disabled_train, get_obj_from_str, + instantiate_from_config, log_txt_as_img) + + +class DiffusionEngine(pl.LightningModule): + def __init__( + self, + network_config, + denoiser_config, + first_stage_config, + conditioner_config: Union[None, Dict, ListConfig, OmegaConf] = None, + sampler_config: Union[None, Dict, ListConfig, OmegaConf] = None, + optimizer_config: Union[None, Dict, ListConfig, OmegaConf] = None, + scheduler_config: Union[None, Dict, ListConfig, OmegaConf] = None, + loss_fn_config: Union[None, Dict, ListConfig, OmegaConf] = None, + network_wrapper: Union[None, str] = None, + ckpt_path: Union[None, str] = None, + use_ema: bool = False, + ema_decay_rate: float = 0.9999, + scale_factor: float = 1.0, + disable_first_stage_autocast=False, + input_key: str = "jpg", + log_keys: Union[List, None] = None, + no_cond_log: bool = False, + compile_model: bool = False, + en_and_decode_n_samples_a_time: Optional[int] = None, + ): + super().__init__() + self.log_keys = log_keys + self.input_key = input_key + self.optimizer_config = default( + optimizer_config, {"target": "torch.optim.AdamW"} + ) + model = instantiate_from_config(network_config) + self.model = get_obj_from_str(default(network_wrapper, OPENAIUNETWRAPPER))( + model, compile_model=compile_model + ) + + self.denoiser = instantiate_from_config(denoiser_config) + self.sampler = ( + instantiate_from_config(sampler_config) + if sampler_config is not None + else None + ) + self.conditioner = instantiate_from_config( + default(conditioner_config, UNCONDITIONAL_CONFIG) + ) + self.scheduler_config = scheduler_config + self._init_first_stage(first_stage_config) + + self.loss_fn = ( + instantiate_from_config(loss_fn_config) + if loss_fn_config is not None + else None + ) + + self.use_ema = use_ema + if self.use_ema: + self.model_ema = LitEma(self.model, decay=ema_decay_rate) + print(f"Keeping EMAs of {len(list(self.model_ema.buffers()))}.") + + self.scale_factor = scale_factor + self.disable_first_stage_autocast = disable_first_stage_autocast + self.no_cond_log = no_cond_log + + if ckpt_path is not None: + self.init_from_ckpt(ckpt_path) + + self.en_and_decode_n_samples_a_time = en_and_decode_n_samples_a_time + + def init_from_ckpt( + self, + path: str, + ) -> None: + if path.endswith("ckpt"): + sd = torch.load(path, map_location="cpu")["state_dict"] + elif path.endswith("safetensors"): + sd = load_safetensors(path) + else: + raise NotImplementedError + + missing, unexpected = self.load_state_dict(sd, strict=False) + print( + f"Restored from {path} with {len(missing)} missing and {len(unexpected)} unexpected keys" + ) + if len(missing) > 0: + print(f"Missing Keys: {missing}") + if len(unexpected) > 0: + print(f"Unexpected Keys: {unexpected}") + + def _init_first_stage(self, config): + model = instantiate_from_config(config).eval() + model.train = disabled_train + for param in model.parameters(): + param.requires_grad = False + self.first_stage_model = model + + def get_input(self, batch): + # assuming unified data format, dataloader returns a dict. + # image tensors should be scaled to -1 ... 1 and in bchw format + return batch[self.input_key] + + @torch.no_grad() + def decode_first_stage(self, z): + z = 1.0 / self.scale_factor * z + n_samples = default(self.en_and_decode_n_samples_a_time, z.shape[0]) + + n_rounds = math.ceil(z.shape[0] / n_samples) + all_out = [] + with torch.autocast("cuda", enabled=not self.disable_first_stage_autocast): + for n in range(n_rounds): + if isinstance(self.first_stage_model.decoder, VideoDecoder): + kwargs = {"timesteps": len(z[n * n_samples : (n + 1) * n_samples])} + else: + kwargs = {} + out = self.first_stage_model.decode( + z[n * n_samples : (n + 1) * n_samples], **kwargs + ) + all_out.append(out) + out = torch.cat(all_out, dim=0) + return out + + @torch.no_grad() + def encode_first_stage(self, x): + n_samples = default(self.en_and_decode_n_samples_a_time, x.shape[0]) + n_rounds = math.ceil(x.shape[0] / n_samples) + all_out = [] + with torch.autocast("cuda", enabled=not self.disable_first_stage_autocast): + for n in range(n_rounds): + out = self.first_stage_model.encode( + x[n * n_samples : (n + 1) * n_samples] + ) + all_out.append(out) + z = torch.cat(all_out, dim=0) + z = self.scale_factor * z + return z + + def forward(self, x, batch): + loss = self.loss_fn(self.model, self.denoiser, self.conditioner, x, batch) + loss_mean = loss.mean() + loss_dict = {"loss": loss_mean} + return loss_mean, loss_dict + + def shared_step(self, batch: Dict) -> Any: + x = self.get_input(batch) + x = self.encode_first_stage(x) + batch["global_step"] = self.global_step + loss, loss_dict = self(x, batch) + return loss, loss_dict + + def training_step(self, batch, batch_idx): + loss, loss_dict = self.shared_step(batch) + + self.log_dict( + loss_dict, prog_bar=True, logger=True, on_step=True, on_epoch=False + ) + + self.log( + "global_step", + self.global_step, + prog_bar=True, + logger=True, + on_step=True, + on_epoch=False, + ) + + if self.scheduler_config is not None: + lr = self.optimizers().param_groups[0]["lr"] + self.log( + "lr_abs", lr, prog_bar=True, logger=True, on_step=True, on_epoch=False + ) + + return loss + + def on_train_start(self, *args, **kwargs): + if self.sampler is None or self.loss_fn is None: + raise ValueError("Sampler and loss function need to be set for training.") + + def on_train_batch_end(self, *args, **kwargs): + if self.use_ema: + self.model_ema(self.model) + + @contextmanager + def ema_scope(self, context=None): + if self.use_ema: + self.model_ema.store(self.model.parameters()) + self.model_ema.copy_to(self.model) + if context is not None: + print(f"{context}: Switched to EMA weights") + try: + yield None + finally: + if self.use_ema: + self.model_ema.restore(self.model.parameters()) + if context is not None: + print(f"{context}: Restored training weights") + + def instantiate_optimizer_from_config(self, params, lr, cfg): + return get_obj_from_str(cfg["target"])( + params, lr=lr, **cfg.get("params", dict()) + ) + + def configure_optimizers(self): + lr = self.learning_rate + params = list(self.model.parameters()) + for embedder in self.conditioner.embedders: + if embedder.is_trainable: + params = params + list(embedder.parameters()) + opt = self.instantiate_optimizer_from_config(params, lr, self.optimizer_config) + if self.scheduler_config is not None: + scheduler = instantiate_from_config(self.scheduler_config) + print("Setting up LambdaLR scheduler...") + scheduler = [ + { + "scheduler": LambdaLR(opt, lr_lambda=scheduler.schedule), + "interval": "step", + "frequency": 1, + } + ] + return [opt], scheduler + return opt + + @torch.no_grad() + def sample( + self, + cond: Dict, + uc: Union[Dict, None] = None, + batch_size: int = 16, + shape: Union[None, Tuple, List] = None, + **kwargs, + ): + randn = torch.randn(batch_size, *shape).to(self.device) + + denoiser = lambda input, sigma, c: self.denoiser( + self.model, input, sigma, c, **kwargs + ) + samples = self.sampler(denoiser, randn, cond, uc=uc) + return samples + + @torch.no_grad() + def log_conditionings(self, batch: Dict, n: int) -> Dict: + """ + Defines heuristics to log different conditionings. + These can be lists of strings (text-to-image), tensors, ints, ... + """ + image_h, image_w = batch[self.input_key].shape[2:] + log = dict() + + for embedder in self.conditioner.embedders: + if ( + (self.log_keys is None) or (embedder.input_key in self.log_keys) + ) and not self.no_cond_log: + x = batch[embedder.input_key][:n] + if isinstance(x, torch.Tensor): + if x.dim() == 1: + # class-conditional, convert integer to string + x = [str(x[i].item()) for i in range(x.shape[0])] + xc = log_txt_as_img((image_h, image_w), x, size=image_h // 4) + elif x.dim() == 2: + # size and crop cond and the like + x = [ + "x".join([str(xx) for xx in x[i].tolist()]) + for i in range(x.shape[0]) + ] + xc = log_txt_as_img((image_h, image_w), x, size=image_h // 20) + else: + raise NotImplementedError() + elif isinstance(x, (List, ListConfig)): + if isinstance(x[0], str): + # strings + xc = log_txt_as_img((image_h, image_w), x, size=image_h // 20) + else: + raise NotImplementedError() + else: + raise NotImplementedError() + log[embedder.input_key] = xc + return log + + @torch.no_grad() + def log_images( + self, + batch: Dict, + N: int = 8, + sample: bool = True, + ucg_keys: List[str] = None, + **kwargs, + ) -> Dict: + conditioner_input_keys = [e.input_key for e in self.conditioner.embedders] + if ucg_keys: + assert all(map(lambda x: x in conditioner_input_keys, ucg_keys)), ( + "Each defined ucg key for sampling must be in the provided conditioner input keys," + f"but we have {ucg_keys} vs. {conditioner_input_keys}" + ) + else: + ucg_keys = conditioner_input_keys + log = dict() + + x = self.get_input(batch) + + c, uc = self.conditioner.get_unconditional_conditioning( + batch, + force_uc_zero_embeddings=ucg_keys + if len(self.conditioner.embedders) > 0 + else [], + ) + + sampling_kwargs = {} + + N = min(x.shape[0], N) + x = x.to(self.device)[:N] + log["inputs"] = x + z = self.encode_first_stage(x) + log["reconstructions"] = self.decode_first_stage(z) + log.update(self.log_conditionings(batch, N)) + + for k in c: + if isinstance(c[k], torch.Tensor): + c[k], uc[k] = map(lambda y: y[k][:N].to(self.device), (c, uc)) + + if sample: + with self.ema_scope("Plotting"): + samples = self.sample( + c, shape=z.shape[1:], uc=uc, batch_size=N, **sampling_kwargs + ) + samples = self.decode_first_stage(samples) + log["samples"] = samples + return log diff --git a/sgm/modules/__init__.py b/sgm/modules/__init__.py new file mode 100644 index 0000000..0db1d77 --- /dev/null +++ b/sgm/modules/__init__.py @@ -0,0 +1,6 @@ +from .encoders.modules import GeneralConditioner + +UNCONDITIONAL_CONFIG = { + "target": "sgm.modules.GeneralConditioner", + "params": {"emb_models": []}, +} diff --git a/sgm/modules/attention.py b/sgm/modules/attention.py new file mode 100644 index 0000000..52a50b7 --- /dev/null +++ b/sgm/modules/attention.py @@ -0,0 +1,759 @@ +import logging +import math +from inspect import isfunction +from typing import Any, Optional + +import torch +import torch.nn.functional as F +from einops import rearrange, repeat +from packaging import version +from torch import nn +from torch.utils.checkpoint import checkpoint + +logpy = logging.getLogger(__name__) + +if version.parse(torch.__version__) >= version.parse("2.0.0"): + SDP_IS_AVAILABLE = True + from torch.backends.cuda import SDPBackend, sdp_kernel + + BACKEND_MAP = { + SDPBackend.MATH: { + "enable_math": True, + "enable_flash": False, + "enable_mem_efficient": False, + }, + SDPBackend.FLASH_ATTENTION: { + "enable_math": False, + "enable_flash": True, + "enable_mem_efficient": False, + }, + SDPBackend.EFFICIENT_ATTENTION: { + "enable_math": False, + "enable_flash": False, + "enable_mem_efficient": True, + }, + None: {"enable_math": True, "enable_flash": True, "enable_mem_efficient": True}, + } +else: + from contextlib import nullcontext + + SDP_IS_AVAILABLE = False + sdp_kernel = nullcontext + BACKEND_MAP = {} + logpy.warn( + f"No SDP backend available, likely because you are running in pytorch " + f"versions < 2.0. In fact, you are using PyTorch {torch.__version__}. " + f"You might want to consider upgrading." + ) + +try: + import xformers + import xformers.ops + + XFORMERS_IS_AVAILABLE = True +except: + XFORMERS_IS_AVAILABLE = False + logpy.warn("no module 'xformers'. Processing without...") + +# from .diffusionmodules.util import mixed_checkpoint as checkpoint + + +def exists(val): + return val is not None + + +def uniq(arr): + return {el: True for el in arr}.keys() + + +def default(val, d): + if exists(val): + return val + return d() if isfunction(d) else d + + +def max_neg_value(t): + return -torch.finfo(t.dtype).max + + +def init_(tensor): + dim = tensor.shape[-1] + std = 1 / math.sqrt(dim) + tensor.uniform_(-std, std) + return tensor + + +# feedforward +class GEGLU(nn.Module): + def __init__(self, dim_in, dim_out): + super().__init__() + self.proj = nn.Linear(dim_in, dim_out * 2) + + def forward(self, x): + x, gate = self.proj(x).chunk(2, dim=-1) + return x * F.gelu(gate) + + +class FeedForward(nn.Module): + def __init__(self, dim, dim_out=None, mult=4, glu=False, dropout=0.0): + super().__init__() + inner_dim = int(dim * mult) + dim_out = default(dim_out, dim) + project_in = ( + nn.Sequential(nn.Linear(dim, inner_dim), nn.GELU()) + if not glu + else GEGLU(dim, inner_dim) + ) + + self.net = nn.Sequential( + project_in, nn.Dropout(dropout), nn.Linear(inner_dim, dim_out) + ) + + def forward(self, x): + return self.net(x) + + +def zero_module(module): + """ + Zero out the parameters of a module and return it. + """ + for p in module.parameters(): + p.detach().zero_() + return module + + +def Normalize(in_channels): + return torch.nn.GroupNorm( + num_groups=32, num_channels=in_channels, eps=1e-6, affine=True + ) + + +class LinearAttention(nn.Module): + def __init__(self, dim, heads=4, dim_head=32): + super().__init__() + self.heads = heads + hidden_dim = dim_head * heads + self.to_qkv = nn.Conv2d(dim, hidden_dim * 3, 1, bias=False) + self.to_out = nn.Conv2d(hidden_dim, dim, 1) + + def forward(self, x): + b, c, h, w = x.shape + qkv = self.to_qkv(x) + q, k, v = rearrange( + qkv, "b (qkv heads c) h w -> qkv b heads c (h w)", heads=self.heads, qkv=3 + ) + k = k.softmax(dim=-1) + context = torch.einsum("bhdn,bhen->bhde", k, v) + out = torch.einsum("bhde,bhdn->bhen", context, q) + out = rearrange( + out, "b heads c (h w) -> b (heads c) h w", heads=self.heads, h=h, w=w + ) + return self.to_out(out) + + +class SelfAttention(nn.Module): + ATTENTION_MODES = ("xformers", "torch", "math") + + def __init__( + self, + dim: int, + num_heads: int = 8, + qkv_bias: bool = False, + qk_scale: Optional[float] = None, + attn_drop: float = 0.0, + proj_drop: float = 0.0, + attn_mode: str = "xformers", + ): + super().__init__() + self.num_heads = num_heads + head_dim = dim // num_heads + self.scale = qk_scale or head_dim**-0.5 + + self.qkv = nn.Linear(dim, dim * 3, bias=qkv_bias) + self.attn_drop = nn.Dropout(attn_drop) + self.proj = nn.Linear(dim, dim) + self.proj_drop = nn.Dropout(proj_drop) + assert attn_mode in self.ATTENTION_MODES + self.attn_mode = attn_mode + + def forward(self, x: torch.Tensor) -> torch.Tensor: + B, L, C = x.shape + + qkv = self.qkv(x) + if self.attn_mode == "torch": + qkv = rearrange( + qkv, "B L (K H D) -> K B H L D", K=3, H=self.num_heads + ).float() + q, k, v = qkv[0], qkv[1], qkv[2] # B H L D + x = torch.nn.functional.scaled_dot_product_attention(q, k, v) + x = rearrange(x, "B H L D -> B L (H D)") + elif self.attn_mode == "xformers": + qkv = rearrange(qkv, "B L (K H D) -> K B L H D", K=3, H=self.num_heads) + q, k, v = qkv[0], qkv[1], qkv[2] # B L H D + x = xformers.ops.memory_efficient_attention(q, k, v) + x = rearrange(x, "B L H D -> B L (H D)", H=self.num_heads) + elif self.attn_mode == "math": + qkv = rearrange(qkv, "B L (K H D) -> K B H L D", K=3, H=self.num_heads) + q, k, v = qkv[0], qkv[1], qkv[2] # B H L D + attn = (q @ k.transpose(-2, -1)) * self.scale + attn = attn.softmax(dim=-1) + attn = self.attn_drop(attn) + x = (attn @ v).transpose(1, 2).reshape(B, L, C) + else: + raise NotImplemented + + x = self.proj(x) + x = self.proj_drop(x) + return x + + +class SpatialSelfAttention(nn.Module): + def __init__(self, in_channels): + super().__init__() + self.in_channels = in_channels + + self.norm = Normalize(in_channels) + self.q = torch.nn.Conv2d( + in_channels, in_channels, kernel_size=1, stride=1, padding=0 + ) + self.k = torch.nn.Conv2d( + in_channels, in_channels, kernel_size=1, stride=1, padding=0 + ) + self.v = torch.nn.Conv2d( + in_channels, in_channels, kernel_size=1, stride=1, padding=0 + ) + self.proj_out = torch.nn.Conv2d( + in_channels, in_channels, kernel_size=1, stride=1, padding=0 + ) + + def forward(self, x): + h_ = x + h_ = self.norm(h_) + q = self.q(h_) + k = self.k(h_) + v = self.v(h_) + + # compute attention + b, c, h, w = q.shape + q = rearrange(q, "b c h w -> b (h w) c") + k = rearrange(k, "b c h w -> b c (h w)") + w_ = torch.einsum("bij,bjk->bik", q, k) + + w_ = w_ * (int(c) ** (-0.5)) + w_ = torch.nn.functional.softmax(w_, dim=2) + + # attend to values + v = rearrange(v, "b c h w -> b c (h w)") + w_ = rearrange(w_, "b i j -> b j i") + h_ = torch.einsum("bij,bjk->bik", v, w_) + h_ = rearrange(h_, "b c (h w) -> b c h w", h=h) + h_ = self.proj_out(h_) + + return x + h_ + + +class CrossAttention(nn.Module): + def __init__( + self, + query_dim, + context_dim=None, + heads=8, + dim_head=64, + dropout=0.0, + backend=None, + ): + super().__init__() + inner_dim = dim_head * heads + context_dim = default(context_dim, query_dim) + + self.scale = dim_head**-0.5 + self.heads = heads + + self.to_q = nn.Linear(query_dim, inner_dim, bias=False) + self.to_k = nn.Linear(context_dim, inner_dim, bias=False) + self.to_v = nn.Linear(context_dim, inner_dim, bias=False) + + self.to_out = nn.Sequential( + nn.Linear(inner_dim, query_dim), nn.Dropout(dropout) + ) + self.backend = backend + + def forward( + self, + x, + context=None, + mask=None, + additional_tokens=None, + n_times_crossframe_attn_in_self=0, + ): + h = self.heads + + if additional_tokens is not None: + # get the number of masked tokens at the beginning of the output sequence + n_tokens_to_mask = additional_tokens.shape[1] + # add additional token + x = torch.cat([additional_tokens, x], dim=1) + + q = self.to_q(x) + context = default(context, x) + k = self.to_k(context) + v = self.to_v(context) + + if n_times_crossframe_attn_in_self: + # reprogramming cross-frame attention as in https://arxiv.org/abs/2303.13439 + assert x.shape[0] % n_times_crossframe_attn_in_self == 0 + n_cp = x.shape[0] // n_times_crossframe_attn_in_self + k = repeat( + k[::n_times_crossframe_attn_in_self], "b ... -> (b n) ...", n=n_cp + ) + v = repeat( + v[::n_times_crossframe_attn_in_self], "b ... -> (b n) ...", n=n_cp + ) + + q, k, v = map(lambda t: rearrange(t, "b n (h d) -> b h n d", h=h), (q, k, v)) + + ## old + """ + sim = einsum('b i d, b j d -> b i j', q, k) * self.scale + del q, k + + if exists(mask): + mask = rearrange(mask, 'b ... -> b (...)') + max_neg_value = -torch.finfo(sim.dtype).max + mask = repeat(mask, 'b j -> (b h) () j', h=h) + sim.masked_fill_(~mask, max_neg_value) + + # attention, what we cannot get enough of + sim = sim.softmax(dim=-1) + + out = einsum('b i j, b j d -> b i d', sim, v) + """ + ## new + with sdp_kernel(**BACKEND_MAP[self.backend]): + # print("dispatching into backend", self.backend, "q/k/v shape: ", q.shape, k.shape, v.shape) + out = F.scaled_dot_product_attention( + q, k, v, attn_mask=mask + ) # scale is dim_head ** -0.5 per default + + del q, k, v + out = rearrange(out, "b h n d -> b n (h d)", h=h) + + if additional_tokens is not None: + # remove additional token + out = out[:, n_tokens_to_mask:] + return self.to_out(out) + + +class MemoryEfficientCrossAttention(nn.Module): + # https://github.com/MatthieuTPHR/diffusers/blob/d80b531ff8060ec1ea982b65a1b8df70f73aa67c/src/diffusers/models/attention.py#L223 + def __init__( + self, query_dim, context_dim=None, heads=8, dim_head=64, dropout=0.0, **kwargs + ): + super().__init__() + logpy.debug( + f"Setting up {self.__class__.__name__}. Query dim is {query_dim}, " + f"context_dim is {context_dim} and using {heads} heads with a " + f"dimension of {dim_head}." + ) + inner_dim = dim_head * heads + context_dim = default(context_dim, query_dim) + + self.heads = heads + self.dim_head = dim_head + + self.to_q = nn.Linear(query_dim, inner_dim, bias=False) + self.to_k = nn.Linear(context_dim, inner_dim, bias=False) + self.to_v = nn.Linear(context_dim, inner_dim, bias=False) + + self.to_out = nn.Sequential( + nn.Linear(inner_dim, query_dim), nn.Dropout(dropout) + ) + self.attention_op: Optional[Any] = None + + def forward( + self, + x, + context=None, + mask=None, + additional_tokens=None, + n_times_crossframe_attn_in_self=0, + ): + if additional_tokens is not None: + # get the number of masked tokens at the beginning of the output sequence + n_tokens_to_mask = additional_tokens.shape[1] + # add additional token + x = torch.cat([additional_tokens, x], dim=1) + q = self.to_q(x) + context = default(context, x) + k = self.to_k(context) + v = self.to_v(context) + + if n_times_crossframe_attn_in_self: + # reprogramming cross-frame attention as in https://arxiv.org/abs/2303.13439 + assert x.shape[0] % n_times_crossframe_attn_in_self == 0 + # n_cp = x.shape[0]//n_times_crossframe_attn_in_self + k = repeat( + k[::n_times_crossframe_attn_in_self], + "b ... -> (b n) ...", + n=n_times_crossframe_attn_in_self, + ) + v = repeat( + v[::n_times_crossframe_attn_in_self], + "b ... -> (b n) ...", + n=n_times_crossframe_attn_in_self, + ) + + b, _, _ = q.shape + q, k, v = map( + lambda t: t.unsqueeze(3) + .reshape(b, t.shape[1], self.heads, self.dim_head) + .permute(0, 2, 1, 3) + .reshape(b * self.heads, t.shape[1], self.dim_head) + .contiguous(), + (q, k, v), + ) + + # actually compute the attention, what we cannot get enough of + if version.parse(xformers.__version__) >= version.parse("0.0.21"): + # NOTE: workaround for + # https://github.com/facebookresearch/xformers/issues/845 + max_bs = 32768 + N = q.shape[0] + n_batches = math.ceil(N / max_bs) + out = list() + for i_batch in range(n_batches): + batch = slice(i_batch * max_bs, (i_batch + 1) * max_bs) + out.append( + xformers.ops.memory_efficient_attention( + q[batch], + k[batch], + v[batch], + attn_bias=None, + op=self.attention_op, + ) + ) + out = torch.cat(out, 0) + else: + out = xformers.ops.memory_efficient_attention( + q, k, v, attn_bias=None, op=self.attention_op + ) + + # TODO: Use this directly in the attention operation, as a bias + if exists(mask): + raise NotImplementedError + out = ( + out.unsqueeze(0) + .reshape(b, self.heads, out.shape[1], self.dim_head) + .permute(0, 2, 1, 3) + .reshape(b, out.shape[1], self.heads * self.dim_head) + ) + if additional_tokens is not None: + # remove additional token + out = out[:, n_tokens_to_mask:] + return self.to_out(out) + + +class BasicTransformerBlock(nn.Module): + ATTENTION_MODES = { + "softmax": CrossAttention, # vanilla attention + "softmax-xformers": MemoryEfficientCrossAttention, # ampere + } + + def __init__( + self, + dim, + n_heads, + d_head, + dropout=0.0, + context_dim=None, + gated_ff=True, + checkpoint=True, + disable_self_attn=False, + attn_mode="softmax", + sdp_backend=None, + ): + super().__init__() + assert attn_mode in self.ATTENTION_MODES + if attn_mode != "softmax" and not XFORMERS_IS_AVAILABLE: + logpy.warn( + f"Attention mode '{attn_mode}' is not available. Falling " + f"back to native attention. This is not a problem in " + f"Pytorch >= 2.0. FYI, you are running with PyTorch " + f"version {torch.__version__}." + ) + attn_mode = "softmax" + elif attn_mode == "softmax" and not SDP_IS_AVAILABLE: + logpy.warn( + "We do not support vanilla attention anymore, as it is too " + "expensive. Sorry." + ) + if not XFORMERS_IS_AVAILABLE: + assert ( + False + ), "Please install xformers via e.g. 'pip install xformers==0.0.16'" + else: + logpy.info("Falling back to xformers efficient attention.") + attn_mode = "softmax-xformers" + attn_cls = self.ATTENTION_MODES[attn_mode] + if version.parse(torch.__version__) >= version.parse("2.0.0"): + assert sdp_backend is None or isinstance(sdp_backend, SDPBackend) + else: + assert sdp_backend is None + self.disable_self_attn = disable_self_attn + self.attn1 = attn_cls( + query_dim=dim, + heads=n_heads, + dim_head=d_head, + dropout=dropout, + context_dim=context_dim if self.disable_self_attn else None, + backend=sdp_backend, + ) # is a self-attention if not self.disable_self_attn + self.ff = FeedForward(dim, dropout=dropout, glu=gated_ff) + self.attn2 = attn_cls( + query_dim=dim, + context_dim=context_dim, + heads=n_heads, + dim_head=d_head, + dropout=dropout, + backend=sdp_backend, + ) # is self-attn if context is none + self.norm1 = nn.LayerNorm(dim) + self.norm2 = nn.LayerNorm(dim) + self.norm3 = nn.LayerNorm(dim) + self.checkpoint = checkpoint + if self.checkpoint: + logpy.debug(f"{self.__class__.__name__} is using checkpointing") + + def forward( + self, x, context=None, additional_tokens=None, n_times_crossframe_attn_in_self=0 + ): + kwargs = {"x": x} + + if context is not None: + kwargs.update({"context": context}) + + if additional_tokens is not None: + kwargs.update({"additional_tokens": additional_tokens}) + + if n_times_crossframe_attn_in_self: + kwargs.update( + {"n_times_crossframe_attn_in_self": n_times_crossframe_attn_in_self} + ) + + # return mixed_checkpoint(self._forward, kwargs, self.parameters(), self.checkpoint) + if self.checkpoint: + # inputs = {"x": x, "context": context} + return checkpoint(self._forward, x, context) + # return checkpoint(self._forward, inputs, self.parameters(), self.checkpoint) + else: + return self._forward(**kwargs) + + def _forward( + self, x, context=None, additional_tokens=None, n_times_crossframe_attn_in_self=0 + ): + x = ( + self.attn1( + self.norm1(x), + context=context if self.disable_self_attn else None, + additional_tokens=additional_tokens, + n_times_crossframe_attn_in_self=n_times_crossframe_attn_in_self + if not self.disable_self_attn + else 0, + ) + + x + ) + x = ( + self.attn2( + self.norm2(x), context=context, additional_tokens=additional_tokens + ) + + x + ) + x = self.ff(self.norm3(x)) + x + return x + + +class BasicTransformerSingleLayerBlock(nn.Module): + ATTENTION_MODES = { + "softmax": CrossAttention, # vanilla attention + "softmax-xformers": MemoryEfficientCrossAttention # on the A100s not quite as fast as the above version + # (todo might depend on head_dim, check, falls back to semi-optimized kernels for dim!=[16,32,64,128]) + } + + def __init__( + self, + dim, + n_heads, + d_head, + dropout=0.0, + context_dim=None, + gated_ff=True, + checkpoint=True, + attn_mode="softmax", + ): + super().__init__() + assert attn_mode in self.ATTENTION_MODES + attn_cls = self.ATTENTION_MODES[attn_mode] + self.attn1 = attn_cls( + query_dim=dim, + heads=n_heads, + dim_head=d_head, + dropout=dropout, + context_dim=context_dim, + ) + self.ff = FeedForward(dim, dropout=dropout, glu=gated_ff) + self.norm1 = nn.LayerNorm(dim) + self.norm2 = nn.LayerNorm(dim) + self.checkpoint = checkpoint + + def forward(self, x, context=None): + # inputs = {"x": x, "context": context} + # return checkpoint(self._forward, inputs, self.parameters(), self.checkpoint) + return checkpoint(self._forward, x, context) + + def _forward(self, x, context=None): + x = self.attn1(self.norm1(x), context=context) + x + x = self.ff(self.norm2(x)) + x + return x + + +class SpatialTransformer(nn.Module): + """ + Transformer block for image-like data. + First, project the input (aka embedding) + and reshape to b, t, d. + Then apply standard transformer action. + Finally, reshape to image + NEW: use_linear for more efficiency instead of the 1x1 convs + """ + + def __init__( + self, + in_channels, + n_heads, + d_head, + depth=1, + dropout=0.0, + context_dim=None, + disable_self_attn=False, + use_linear=False, + attn_type="softmax", + use_checkpoint=True, + # sdp_backend=SDPBackend.FLASH_ATTENTION + sdp_backend=None, + ): + super().__init__() + logpy.debug( + f"constructing {self.__class__.__name__} of depth {depth} w/ " + f"{in_channels} channels and {n_heads} heads." + ) + + if exists(context_dim) and not isinstance(context_dim, list): + context_dim = [context_dim] + if exists(context_dim) and isinstance(context_dim, list): + if depth != len(context_dim): + logpy.warn( + f"{self.__class__.__name__}: Found context dims " + f"{context_dim} of depth {len(context_dim)}, which does not " + f"match the specified 'depth' of {depth}. Setting context_dim " + f"to {depth * [context_dim[0]]} now." + ) + # depth does not match context dims. + assert all( + map(lambda x: x == context_dim[0], context_dim) + ), "need homogenous context_dim to match depth automatically" + context_dim = depth * [context_dim[0]] + elif context_dim is None: + context_dim = [None] * depth + self.in_channels = in_channels + inner_dim = n_heads * d_head + self.norm = Normalize(in_channels) + if not use_linear: + self.proj_in = nn.Conv2d( + in_channels, inner_dim, kernel_size=1, stride=1, padding=0 + ) + else: + self.proj_in = nn.Linear(in_channels, inner_dim) + + self.transformer_blocks = nn.ModuleList( + [ + BasicTransformerBlock( + inner_dim, + n_heads, + d_head, + dropout=dropout, + context_dim=context_dim[d], + disable_self_attn=disable_self_attn, + attn_mode=attn_type, + checkpoint=use_checkpoint, + sdp_backend=sdp_backend, + ) + for d in range(depth) + ] + ) + if not use_linear: + self.proj_out = zero_module( + nn.Conv2d(inner_dim, in_channels, kernel_size=1, stride=1, padding=0) + ) + else: + # self.proj_out = zero_module(nn.Linear(in_channels, inner_dim)) + self.proj_out = zero_module(nn.Linear(inner_dim, in_channels)) + self.use_linear = use_linear + + def forward(self, x, context=None): + # note: if no context is given, cross-attention defaults to self-attention + if not isinstance(context, list): + context = [context] + b, c, h, w = x.shape + x_in = x + x = self.norm(x) + if not self.use_linear: + x = self.proj_in(x) + x = rearrange(x, "b c h w -> b (h w) c").contiguous() + if self.use_linear: + x = self.proj_in(x) + for i, block in enumerate(self.transformer_blocks): + if i > 0 and len(context) == 1: + i = 0 # use same context for each block + x = block(x, context=context[i]) + if self.use_linear: + x = self.proj_out(x) + x = rearrange(x, "b (h w) c -> b c h w", h=h, w=w).contiguous() + if not self.use_linear: + x = self.proj_out(x) + return x + x_in + + +class SimpleTransformer(nn.Module): + def __init__( + self, + dim: int, + depth: int, + heads: int, + dim_head: int, + context_dim: Optional[int] = None, + dropout: float = 0.0, + checkpoint: bool = True, + ): + super().__init__() + self.layers = nn.ModuleList([]) + for _ in range(depth): + self.layers.append( + BasicTransformerBlock( + dim, + heads, + dim_head, + dropout=dropout, + context_dim=context_dim, + attn_mode="softmax-xformers", + checkpoint=checkpoint, + ) + ) + + def forward( + self, + x: torch.Tensor, + context: Optional[torch.Tensor] = None, + ) -> torch.Tensor: + for layer in self.layers: + x = layer(x, context) + return x diff --git a/sgm/modules/autoencoding/__init__.py b/sgm/modules/autoencoding/__init__.py new file mode 100644 index 0000000..e69de29 diff --git a/sgm/modules/autoencoding/losses/__init__.py b/sgm/modules/autoencoding/losses/__init__.py new file mode 100644 index 0000000..6b316c7 --- /dev/null +++ b/sgm/modules/autoencoding/losses/__init__.py @@ -0,0 +1,7 @@ +__all__ = [ + "GeneralLPIPSWithDiscriminator", + "LatentLPIPS", +] + +from .discriminator_loss import GeneralLPIPSWithDiscriminator +from .lpips import LatentLPIPS diff --git a/sgm/modules/autoencoding/losses/discriminator_loss.py b/sgm/modules/autoencoding/losses/discriminator_loss.py new file mode 100644 index 0000000..09b6829 --- /dev/null +++ b/sgm/modules/autoencoding/losses/discriminator_loss.py @@ -0,0 +1,306 @@ +from typing import Dict, Iterator, List, Optional, Tuple, Union + +import numpy as np +import torch +import torch.nn as nn +import torchvision +from einops import rearrange +from matplotlib import colormaps +from matplotlib import pyplot as plt + +from ....util import default, instantiate_from_config +from ..lpips.loss.lpips import LPIPS +from ..lpips.model.model import weights_init +from ..lpips.vqperceptual import hinge_d_loss, vanilla_d_loss + + +class GeneralLPIPSWithDiscriminator(nn.Module): + def __init__( + self, + disc_start: int, + logvar_init: float = 0.0, + disc_num_layers: int = 3, + disc_in_channels: int = 3, + disc_factor: float = 1.0, + disc_weight: float = 1.0, + perceptual_weight: float = 1.0, + disc_loss: str = "hinge", + scale_input_to_tgt_size: bool = False, + dims: int = 2, + learn_logvar: bool = False, + regularization_weights: Union[None, Dict[str, float]] = None, + additional_log_keys: Optional[List[str]] = None, + discriminator_config: Optional[Dict] = None, + ): + super().__init__() + self.dims = dims + if self.dims > 2: + print( + f"running with dims={dims}. This means that for perceptual loss " + f"calculation, the LPIPS loss will be applied to each frame " + f"independently." + ) + self.scale_input_to_tgt_size = scale_input_to_tgt_size + assert disc_loss in ["hinge", "vanilla"] + self.perceptual_loss = LPIPS().eval() + self.perceptual_weight = perceptual_weight + # output log variance + self.logvar = nn.Parameter( + torch.full((), logvar_init), requires_grad=learn_logvar + ) + self.learn_logvar = learn_logvar + + discriminator_config = default( + discriminator_config, + { + "target": "sgm.modules.autoencoding.lpips.model.model.NLayerDiscriminator", + "params": { + "input_nc": disc_in_channels, + "n_layers": disc_num_layers, + "use_actnorm": False, + }, + }, + ) + + self.discriminator = instantiate_from_config(discriminator_config).apply( + weights_init + ) + self.discriminator_iter_start = disc_start + self.disc_loss = hinge_d_loss if disc_loss == "hinge" else vanilla_d_loss + self.disc_factor = disc_factor + self.discriminator_weight = disc_weight + self.regularization_weights = default(regularization_weights, {}) + + self.forward_keys = [ + "optimizer_idx", + "global_step", + "last_layer", + "split", + "regularization_log", + ] + + self.additional_log_keys = set(default(additional_log_keys, [])) + self.additional_log_keys.update(set(self.regularization_weights.keys())) + + def get_trainable_parameters(self) -> Iterator[nn.Parameter]: + return self.discriminator.parameters() + + def get_trainable_autoencoder_parameters(self) -> Iterator[nn.Parameter]: + if self.learn_logvar: + yield self.logvar + yield from () + + @torch.no_grad() + def log_images( + self, inputs: torch.Tensor, reconstructions: torch.Tensor + ) -> Dict[str, torch.Tensor]: + # calc logits of real/fake + logits_real = self.discriminator(inputs.contiguous().detach()) + if len(logits_real.shape) < 4: + # Non patch-discriminator + return dict() + logits_fake = self.discriminator(reconstructions.contiguous().detach()) + # -> (b, 1, h, w) + + # parameters for colormapping + high = max(logits_fake.abs().max(), logits_real.abs().max()).item() + cmap = colormaps["PiYG"] # diverging colormap + + def to_colormap(logits: torch.Tensor) -> torch.Tensor: + """(b, 1, ...) -> (b, 3, ...)""" + logits = (logits + high) / (2 * high) + logits_np = cmap(logits.cpu().numpy())[..., :3] # truncate alpha channel + # -> (b, 1, ..., 3) + logits = torch.from_numpy(logits_np).to(logits.device) + return rearrange(logits, "b 1 ... c -> b c ...") + + logits_real = torch.nn.functional.interpolate( + logits_real, + size=inputs.shape[-2:], + mode="nearest", + antialias=False, + ) + logits_fake = torch.nn.functional.interpolate( + logits_fake, + size=reconstructions.shape[-2:], + mode="nearest", + antialias=False, + ) + + # alpha value of logits for overlay + alpha_real = torch.abs(logits_real) / high + alpha_fake = torch.abs(logits_fake) / high + # -> (b, 1, h, w) in range [0, 0.5] + # alpha value of lines don't really matter, since the values are the same + # for both images and logits anyway + grid_alpha_real = torchvision.utils.make_grid(alpha_real, nrow=4) + grid_alpha_fake = torchvision.utils.make_grid(alpha_fake, nrow=4) + grid_alpha = 0.8 * torch.cat((grid_alpha_real, grid_alpha_fake), dim=1) + # -> (1, h, w) + # blend logits and images together + + # prepare logits for plotting + logits_real = to_colormap(logits_real) + logits_fake = to_colormap(logits_fake) + # resize logits + # -> (b, 3, h, w) + + # make some grids + # add all logits to one plot + logits_real = torchvision.utils.make_grid(logits_real, nrow=4) + logits_fake = torchvision.utils.make_grid(logits_fake, nrow=4) + # I just love how torchvision calls the number of columns `nrow` + grid_logits = torch.cat((logits_real, logits_fake), dim=1) + # -> (3, h, w) + + grid_images_real = torchvision.utils.make_grid(0.5 * inputs + 0.5, nrow=4) + grid_images_fake = torchvision.utils.make_grid( + 0.5 * reconstructions + 0.5, nrow=4 + ) + grid_images = torch.cat((grid_images_real, grid_images_fake), dim=1) + # -> (3, h, w) in range [0, 1] + + grid_blend = grid_alpha * grid_logits + (1 - grid_alpha) * grid_images + + # Create labeled colorbar + dpi = 100 + height = 128 / dpi + width = grid_logits.shape[2] / dpi + fig, ax = plt.subplots(figsize=(width, height), dpi=dpi) + img = ax.imshow(np.array([[-high, high]]), cmap=cmap) + plt.colorbar( + img, + cax=ax, + orientation="horizontal", + fraction=0.9, + aspect=width / height, + pad=0.0, + ) + img.set_visible(False) + fig.tight_layout() + fig.canvas.draw() + # manually convert figure to numpy + cbar_np = np.frombuffer(fig.canvas.tostring_rgb(), dtype=np.uint8) + cbar_np = cbar_np.reshape(fig.canvas.get_width_height()[::-1] + (3,)) + cbar = torch.from_numpy(cbar_np.copy()).to(grid_logits.dtype) / 255.0 + cbar = rearrange(cbar, "h w c -> c h w").to(grid_logits.device) + + # Add colorbar to plot + annotated_grid = torch.cat((grid_logits, cbar), dim=1) + blended_grid = torch.cat((grid_blend, cbar), dim=1) + return { + "vis_logits": 2 * annotated_grid[None, ...] - 1, + "vis_logits_blended": 2 * blended_grid[None, ...] - 1, + } + + def calculate_adaptive_weight( + self, nll_loss: torch.Tensor, g_loss: torch.Tensor, last_layer: torch.Tensor + ) -> torch.Tensor: + nll_grads = torch.autograd.grad(nll_loss, last_layer, retain_graph=True)[0] + g_grads = torch.autograd.grad(g_loss, last_layer, retain_graph=True)[0] + + d_weight = torch.norm(nll_grads) / (torch.norm(g_grads) + 1e-4) + d_weight = torch.clamp(d_weight, 0.0, 1e4).detach() + d_weight = d_weight * self.discriminator_weight + return d_weight + + def forward( + self, + inputs: torch.Tensor, + reconstructions: torch.Tensor, + *, # added because I changed the order here + regularization_log: Dict[str, torch.Tensor], + optimizer_idx: int, + global_step: int, + last_layer: torch.Tensor, + split: str = "train", + weights: Union[None, float, torch.Tensor] = None, + ) -> Tuple[torch.Tensor, dict]: + if self.scale_input_to_tgt_size: + inputs = torch.nn.functional.interpolate( + inputs, reconstructions.shape[2:], mode="bicubic", antialias=True + ) + + if self.dims > 2: + inputs, reconstructions = map( + lambda x: rearrange(x, "b c t h w -> (b t) c h w"), + (inputs, reconstructions), + ) + + rec_loss = torch.abs(inputs.contiguous() - reconstructions.contiguous()) + if self.perceptual_weight > 0: + p_loss = self.perceptual_loss( + inputs.contiguous(), reconstructions.contiguous() + ) + rec_loss = rec_loss + self.perceptual_weight * p_loss + + nll_loss, weighted_nll_loss = self.get_nll_loss(rec_loss, weights) + + # now the GAN part + if optimizer_idx == 0: + # generator update + if global_step >= self.discriminator_iter_start or not self.training: + logits_fake = self.discriminator(reconstructions.contiguous()) + g_loss = -torch.mean(logits_fake) + if self.training: + d_weight = self.calculate_adaptive_weight( + nll_loss, g_loss, last_layer=last_layer + ) + else: + d_weight = torch.tensor(1.0) + else: + d_weight = torch.tensor(0.0) + g_loss = torch.tensor(0.0, requires_grad=True) + + loss = weighted_nll_loss + d_weight * self.disc_factor * g_loss + log = dict() + for k in regularization_log: + if k in self.regularization_weights: + loss = loss + self.regularization_weights[k] * regularization_log[k] + if k in self.additional_log_keys: + log[f"{split}/{k}"] = regularization_log[k].detach().float().mean() + + log.update( + { + f"{split}/loss/total": loss.clone().detach().mean(), + f"{split}/loss/nll": nll_loss.detach().mean(), + f"{split}/loss/rec": rec_loss.detach().mean(), + f"{split}/loss/g": g_loss.detach().mean(), + f"{split}/scalars/logvar": self.logvar.detach(), + f"{split}/scalars/d_weight": d_weight.detach(), + } + ) + + return loss, log + elif optimizer_idx == 1: + # second pass for discriminator update + logits_real = self.discriminator(inputs.contiguous().detach()) + logits_fake = self.discriminator(reconstructions.contiguous().detach()) + + if global_step >= self.discriminator_iter_start or not self.training: + d_loss = self.disc_factor * self.disc_loss(logits_real, logits_fake) + else: + d_loss = torch.tensor(0.0, requires_grad=True) + + log = { + f"{split}/loss/disc": d_loss.clone().detach().mean(), + f"{split}/logits/real": logits_real.detach().mean(), + f"{split}/logits/fake": logits_fake.detach().mean(), + } + return d_loss, log + else: + raise NotImplementedError(f"Unknown optimizer_idx {optimizer_idx}") + + def get_nll_loss( + self, + rec_loss: torch.Tensor, + weights: Optional[Union[float, torch.Tensor]] = None, + ) -> Tuple[torch.Tensor, torch.Tensor]: + nll_loss = rec_loss / torch.exp(self.logvar) + self.logvar + weighted_nll_loss = nll_loss + if weights is not None: + weighted_nll_loss = weights * nll_loss + weighted_nll_loss = torch.sum(weighted_nll_loss) / weighted_nll_loss.shape[0] + nll_loss = torch.sum(nll_loss) / nll_loss.shape[0] + + return nll_loss, weighted_nll_loss diff --git a/sgm/modules/autoencoding/losses/lpips.py b/sgm/modules/autoencoding/losses/lpips.py new file mode 100644 index 0000000..b329fcc --- /dev/null +++ b/sgm/modules/autoencoding/losses/lpips.py @@ -0,0 +1,73 @@ +import torch +import torch.nn as nn + +from ....util import default, instantiate_from_config +from ..lpips.loss.lpips import LPIPS + + +class LatentLPIPS(nn.Module): + def __init__( + self, + decoder_config, + perceptual_weight=1.0, + latent_weight=1.0, + scale_input_to_tgt_size=False, + scale_tgt_to_input_size=False, + perceptual_weight_on_inputs=0.0, + ): + super().__init__() + self.scale_input_to_tgt_size = scale_input_to_tgt_size + self.scale_tgt_to_input_size = scale_tgt_to_input_size + self.init_decoder(decoder_config) + self.perceptual_loss = LPIPS().eval() + self.perceptual_weight = perceptual_weight + self.latent_weight = latent_weight + self.perceptual_weight_on_inputs = perceptual_weight_on_inputs + + def init_decoder(self, config): + self.decoder = instantiate_from_config(config) + if hasattr(self.decoder, "encoder"): + del self.decoder.encoder + + def forward(self, latent_inputs, latent_predictions, image_inputs, split="train"): + log = dict() + loss = (latent_inputs - latent_predictions) ** 2 + log[f"{split}/latent_l2_loss"] = loss.mean().detach() + image_reconstructions = None + if self.perceptual_weight > 0.0: + image_reconstructions = self.decoder.decode(latent_predictions) + image_targets = self.decoder.decode(latent_inputs) + perceptual_loss = self.perceptual_loss( + image_targets.contiguous(), image_reconstructions.contiguous() + ) + loss = ( + self.latent_weight * loss.mean() + + self.perceptual_weight * perceptual_loss.mean() + ) + log[f"{split}/perceptual_loss"] = perceptual_loss.mean().detach() + + if self.perceptual_weight_on_inputs > 0.0: + image_reconstructions = default( + image_reconstructions, self.decoder.decode(latent_predictions) + ) + if self.scale_input_to_tgt_size: + image_inputs = torch.nn.functional.interpolate( + image_inputs, + image_reconstructions.shape[2:], + mode="bicubic", + antialias=True, + ) + elif self.scale_tgt_to_input_size: + image_reconstructions = torch.nn.functional.interpolate( + image_reconstructions, + image_inputs.shape[2:], + mode="bicubic", + antialias=True, + ) + + perceptual_loss2 = self.perceptual_loss( + image_inputs.contiguous(), image_reconstructions.contiguous() + ) + loss = loss + self.perceptual_weight_on_inputs * perceptual_loss2.mean() + log[f"{split}/perceptual_loss_on_inputs"] = perceptual_loss2.mean().detach() + return loss, log diff --git a/sgm/modules/autoencoding/lpips/__init__.py b/sgm/modules/autoencoding/lpips/__init__.py new file mode 100644 index 0000000..e69de29 diff --git a/sgm/modules/autoencoding/lpips/loss/.gitignore b/sgm/modules/autoencoding/lpips/loss/.gitignore new file mode 100644 index 0000000..a92958a --- /dev/null +++ b/sgm/modules/autoencoding/lpips/loss/.gitignore @@ -0,0 +1 @@ +vgg.pth \ No newline at end of file diff --git a/sgm/modules/autoencoding/lpips/loss/LICENSE b/sgm/modules/autoencoding/lpips/loss/LICENSE new file mode 100644 index 0000000..924cfc8 --- /dev/null +++ b/sgm/modules/autoencoding/lpips/loss/LICENSE @@ -0,0 +1,23 @@ +Copyright (c) 2018, Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, Oliver Wang +All rights reserved. + +Redistribution and use in source and binary forms, with or without +modification, are permitted provided that the following conditions are met: + +* Redistributions of source code must retain the above copyright notice, this + list of conditions and the following disclaimer. + +* Redistributions in binary form must reproduce the above copyright notice, + this list of conditions and the following disclaimer in the documentation + and/or other materials provided with the distribution. + +THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" +AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE +IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE +DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE +FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL +DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR +SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER +CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, +OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE +OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. \ No newline at end of file diff --git a/sgm/modules/autoencoding/lpips/loss/__init__.py b/sgm/modules/autoencoding/lpips/loss/__init__.py new file mode 100644 index 0000000..e69de29 diff --git a/sgm/modules/autoencoding/lpips/loss/lpips.py b/sgm/modules/autoencoding/lpips/loss/lpips.py new file mode 100644 index 0000000..3e34f3d --- /dev/null +++ b/sgm/modules/autoencoding/lpips/loss/lpips.py @@ -0,0 +1,147 @@ +"""Stripped version of https://github.com/richzhang/PerceptualSimilarity/tree/master/models""" + +from collections import namedtuple + +import torch +import torch.nn as nn +from torchvision import models + +from ..util import get_ckpt_path + + +class LPIPS(nn.Module): + # Learned perceptual metric + def __init__(self, use_dropout=True): + super().__init__() + self.scaling_layer = ScalingLayer() + self.chns = [64, 128, 256, 512, 512] # vg16 features + self.net = vgg16(pretrained=True, requires_grad=False) + self.lin0 = NetLinLayer(self.chns[0], use_dropout=use_dropout) + self.lin1 = NetLinLayer(self.chns[1], use_dropout=use_dropout) + self.lin2 = NetLinLayer(self.chns[2], use_dropout=use_dropout) + self.lin3 = NetLinLayer(self.chns[3], use_dropout=use_dropout) + self.lin4 = NetLinLayer(self.chns[4], use_dropout=use_dropout) + self.load_from_pretrained() + for param in self.parameters(): + param.requires_grad = False + + def load_from_pretrained(self, name="vgg_lpips"): + ckpt = get_ckpt_path(name, "sgm/modules/autoencoding/lpips/loss") + self.load_state_dict( + torch.load(ckpt, map_location=torch.device("cpu")), strict=False + ) + print("loaded pretrained LPIPS loss from {}".format(ckpt)) + + @classmethod + def from_pretrained(cls, name="vgg_lpips"): + if name != "vgg_lpips": + raise NotImplementedError + model = cls() + ckpt = get_ckpt_path(name) + model.load_state_dict( + torch.load(ckpt, map_location=torch.device("cpu")), strict=False + ) + return model + + def forward(self, input, target): + in0_input, in1_input = (self.scaling_layer(input), self.scaling_layer(target)) + outs0, outs1 = self.net(in0_input), self.net(in1_input) + feats0, feats1, diffs = {}, {}, {} + lins = [self.lin0, self.lin1, self.lin2, self.lin3, self.lin4] + for kk in range(len(self.chns)): + feats0[kk], feats1[kk] = normalize_tensor(outs0[kk]), normalize_tensor( + outs1[kk] + ) + diffs[kk] = (feats0[kk] - feats1[kk]) ** 2 + + res = [ + spatial_average(lins[kk].model(diffs[kk]), keepdim=True) + for kk in range(len(self.chns)) + ] + val = res[0] + for l in range(1, len(self.chns)): + val += res[l] + return val + + +class ScalingLayer(nn.Module): + def __init__(self): + super(ScalingLayer, self).__init__() + self.register_buffer( + "shift", torch.Tensor([-0.030, -0.088, -0.188])[None, :, None, None] + ) + self.register_buffer( + "scale", torch.Tensor([0.458, 0.448, 0.450])[None, :, None, None] + ) + + def forward(self, inp): + return (inp - self.shift) / self.scale + + +class NetLinLayer(nn.Module): + """A single linear layer which does a 1x1 conv""" + + def __init__(self, chn_in, chn_out=1, use_dropout=False): + super(NetLinLayer, self).__init__() + layers = ( + [ + nn.Dropout(), + ] + if (use_dropout) + else [] + ) + layers += [ + nn.Conv2d(chn_in, chn_out, 1, stride=1, padding=0, bias=False), + ] + self.model = nn.Sequential(*layers) + + +class vgg16(torch.nn.Module): + def __init__(self, requires_grad=False, pretrained=True): + super(vgg16, self).__init__() + vgg_pretrained_features = models.vgg16(pretrained=pretrained).features + self.slice1 = torch.nn.Sequential() + self.slice2 = torch.nn.Sequential() + self.slice3 = torch.nn.Sequential() + self.slice4 = torch.nn.Sequential() + self.slice5 = torch.nn.Sequential() + self.N_slices = 5 + for x in range(4): + self.slice1.add_module(str(x), vgg_pretrained_features[x]) + for x in range(4, 9): + self.slice2.add_module(str(x), vgg_pretrained_features[x]) + for x in range(9, 16): + self.slice3.add_module(str(x), vgg_pretrained_features[x]) + for x in range(16, 23): + self.slice4.add_module(str(x), vgg_pretrained_features[x]) + for x in range(23, 30): + self.slice5.add_module(str(x), vgg_pretrained_features[x]) + if not requires_grad: + for param in self.parameters(): + param.requires_grad = False + + def forward(self, X): + h = self.slice1(X) + h_relu1_2 = h + h = self.slice2(h) + h_relu2_2 = h + h = self.slice3(h) + h_relu3_3 = h + h = self.slice4(h) + h_relu4_3 = h + h = self.slice5(h) + h_relu5_3 = h + vgg_outputs = namedtuple( + "VggOutputs", ["relu1_2", "relu2_2", "relu3_3", "relu4_3", "relu5_3"] + ) + out = vgg_outputs(h_relu1_2, h_relu2_2, h_relu3_3, h_relu4_3, h_relu5_3) + return out + + +def normalize_tensor(x, eps=1e-10): + norm_factor = torch.sqrt(torch.sum(x**2, dim=1, keepdim=True)) + return x / (norm_factor + eps) + + +def spatial_average(x, keepdim=True): + return x.mean([2, 3], keepdim=keepdim) diff --git a/sgm/modules/autoencoding/lpips/model/LICENSE b/sgm/modules/autoencoding/lpips/model/LICENSE new file mode 100644 index 0000000..4b356e6 --- /dev/null +++ b/sgm/modules/autoencoding/lpips/model/LICENSE @@ -0,0 +1,58 @@ +Copyright (c) 2017, Jun-Yan Zhu and Taesung Park +All rights reserved. + +Redistribution and use in source and binary forms, with or without +modification, are permitted provided that the following conditions are met: + +* Redistributions of source code must retain the above copyright notice, this + list of conditions and the following disclaimer. + +* Redistributions in binary form must reproduce the above copyright notice, + this list of conditions and the following disclaimer in the documentation + and/or other materials provided with the distribution. + +THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" +AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE +IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE +DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE +FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL +DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR +SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER +CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, +OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE +OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. + + +--------------------------- LICENSE FOR pix2pix -------------------------------- +BSD License + +For pix2pix software +Copyright (c) 2016, Phillip Isola and Jun-Yan Zhu +All rights reserved. + +Redistribution and use in source and binary forms, with or without +modification, are permitted provided that the following conditions are met: + +* Redistributions of source code must retain the above copyright notice, this + list of conditions and the following disclaimer. + +* Redistributions in binary form must reproduce the above copyright notice, + this list of conditions and the following disclaimer in the documentation + and/or other materials provided with the distribution. + +----------------------------- LICENSE FOR DCGAN -------------------------------- +BSD License + +For dcgan.torch software + +Copyright (c) 2015, Facebook, Inc. All rights reserved. + +Redistribution and use in source and binary forms, with or without modification, are permitted provided that the following conditions are met: + +Redistributions of source code must retain the above copyright notice, this list of conditions and the following disclaimer. + +Redistributions in binary form must reproduce the above copyright notice, this list of conditions and the following disclaimer in the documentation and/or other materials provided with the distribution. + +Neither the name Facebook nor the names of its contributors may be used to endorse or promote products derived from this software without specific prior written permission. + +THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. \ No newline at end of file diff --git a/sgm/modules/autoencoding/lpips/model/__init__.py b/sgm/modules/autoencoding/lpips/model/__init__.py new file mode 100644 index 0000000..e69de29 diff --git a/sgm/modules/autoencoding/lpips/model/model.py b/sgm/modules/autoencoding/lpips/model/model.py new file mode 100644 index 0000000..66357d4 --- /dev/null +++ b/sgm/modules/autoencoding/lpips/model/model.py @@ -0,0 +1,88 @@ +import functools + +import torch.nn as nn + +from ..util import ActNorm + + +def weights_init(m): + classname = m.__class__.__name__ + if classname.find("Conv") != -1: + nn.init.normal_(m.weight.data, 0.0, 0.02) + elif classname.find("BatchNorm") != -1: + nn.init.normal_(m.weight.data, 1.0, 0.02) + nn.init.constant_(m.bias.data, 0) + + +class NLayerDiscriminator(nn.Module): + """Defines a PatchGAN discriminator as in Pix2Pix + --> see https://github.com/junyanz/pytorch-CycleGAN-and-pix2pix/blob/master/models/networks.py + """ + + def __init__(self, input_nc=3, ndf=64, n_layers=3, use_actnorm=False): + """Construct a PatchGAN discriminator + Parameters: + input_nc (int) -- the number of channels in input images + ndf (int) -- the number of filters in the last conv layer + n_layers (int) -- the number of conv layers in the discriminator + norm_layer -- normalization layer + """ + super(NLayerDiscriminator, self).__init__() + if not use_actnorm: + norm_layer = nn.BatchNorm2d + else: + norm_layer = ActNorm + if ( + type(norm_layer) == functools.partial + ): # no need to use bias as BatchNorm2d has affine parameters + use_bias = norm_layer.func != nn.BatchNorm2d + else: + use_bias = norm_layer != nn.BatchNorm2d + + kw = 4 + padw = 1 + sequence = [ + nn.Conv2d(input_nc, ndf, kernel_size=kw, stride=2, padding=padw), + nn.LeakyReLU(0.2, True), + ] + nf_mult = 1 + nf_mult_prev = 1 + for n in range(1, n_layers): # gradually increase the number of filters + nf_mult_prev = nf_mult + nf_mult = min(2**n, 8) + sequence += [ + nn.Conv2d( + ndf * nf_mult_prev, + ndf * nf_mult, + kernel_size=kw, + stride=2, + padding=padw, + bias=use_bias, + ), + norm_layer(ndf * nf_mult), + nn.LeakyReLU(0.2, True), + ] + + nf_mult_prev = nf_mult + nf_mult = min(2**n_layers, 8) + sequence += [ + nn.Conv2d( + ndf * nf_mult_prev, + ndf * nf_mult, + kernel_size=kw, + stride=1, + padding=padw, + bias=use_bias, + ), + norm_layer(ndf * nf_mult), + nn.LeakyReLU(0.2, True), + ] + + sequence += [ + nn.Conv2d(ndf * nf_mult, 1, kernel_size=kw, stride=1, padding=padw) + ] # output 1 channel prediction map + self.main = nn.Sequential(*sequence) + + def forward(self, input): + """Standard forward.""" + return self.main(input) diff --git a/sgm/modules/autoencoding/lpips/util.py b/sgm/modules/autoencoding/lpips/util.py new file mode 100644 index 0000000..49c76e3 --- /dev/null +++ b/sgm/modules/autoencoding/lpips/util.py @@ -0,0 +1,128 @@ +import hashlib +import os + +import requests +import torch +import torch.nn as nn +from tqdm import tqdm + +URL_MAP = {"vgg_lpips": "https://heibox.uni-heidelberg.de/f/607503859c864bc1b30b/?dl=1"} + +CKPT_MAP = {"vgg_lpips": "vgg.pth"} + +MD5_MAP = {"vgg_lpips": "d507d7349b931f0638a25a48a722f98a"} + + +def download(url, local_path, chunk_size=1024): + os.makedirs(os.path.split(local_path)[0], exist_ok=True) + with requests.get(url, stream=True) as r: + total_size = int(r.headers.get("content-length", 0)) + with tqdm(total=total_size, unit="B", unit_scale=True) as pbar: + with open(local_path, "wb") as f: + for data in r.iter_content(chunk_size=chunk_size): + if data: + f.write(data) + pbar.update(chunk_size) + + +def md5_hash(path): + with open(path, "rb") as f: + content = f.read() + return hashlib.md5(content).hexdigest() + + +def get_ckpt_path(name, root, check=False): + assert name in URL_MAP + path = os.path.join(root, CKPT_MAP[name]) + if not os.path.exists(path) or (check and not md5_hash(path) == MD5_MAP[name]): + print("Downloading {} model from {} to {}".format(name, URL_MAP[name], path)) + download(URL_MAP[name], path) + md5 = md5_hash(path) + assert md5 == MD5_MAP[name], md5 + return path + + +class ActNorm(nn.Module): + def __init__( + self, num_features, logdet=False, affine=True, allow_reverse_init=False + ): + assert affine + super().__init__() + self.logdet = logdet + self.loc = nn.Parameter(torch.zeros(1, num_features, 1, 1)) + self.scale = nn.Parameter(torch.ones(1, num_features, 1, 1)) + self.allow_reverse_init = allow_reverse_init + + self.register_buffer("initialized", torch.tensor(0, dtype=torch.uint8)) + + def initialize(self, input): + with torch.no_grad(): + flatten = input.permute(1, 0, 2, 3).contiguous().view(input.shape[1], -1) + mean = ( + flatten.mean(1) + .unsqueeze(1) + .unsqueeze(2) + .unsqueeze(3) + .permute(1, 0, 2, 3) + ) + std = ( + flatten.std(1) + .unsqueeze(1) + .unsqueeze(2) + .unsqueeze(3) + .permute(1, 0, 2, 3) + ) + + self.loc.data.copy_(-mean) + self.scale.data.copy_(1 / (std + 1e-6)) + + def forward(self, input, reverse=False): + if reverse: + return self.reverse(input) + if len(input.shape) == 2: + input = input[:, :, None, None] + squeeze = True + else: + squeeze = False + + _, _, height, width = input.shape + + if self.training and self.initialized.item() == 0: + self.initialize(input) + self.initialized.fill_(1) + + h = self.scale * (input + self.loc) + + if squeeze: + h = h.squeeze(-1).squeeze(-1) + + if self.logdet: + log_abs = torch.log(torch.abs(self.scale)) + logdet = height * width * torch.sum(log_abs) + logdet = logdet * torch.ones(input.shape[0]).to(input) + return h, logdet + + return h + + def reverse(self, output): + if self.training and self.initialized.item() == 0: + if not self.allow_reverse_init: + raise RuntimeError( + "Initializing ActNorm in reverse direction is " + "disabled by default. Use allow_reverse_init=True to enable." + ) + else: + self.initialize(output) + self.initialized.fill_(1) + + if len(output.shape) == 2: + output = output[:, :, None, None] + squeeze = True + else: + squeeze = False + + h = output / self.scale - self.loc + + if squeeze: + h = h.squeeze(-1).squeeze(-1) + return h diff --git a/sgm/modules/autoencoding/lpips/vqperceptual.py b/sgm/modules/autoencoding/lpips/vqperceptual.py new file mode 100644 index 0000000..6195f0a --- /dev/null +++ b/sgm/modules/autoencoding/lpips/vqperceptual.py @@ -0,0 +1,17 @@ +import torch +import torch.nn.functional as F + + +def hinge_d_loss(logits_real, logits_fake): + loss_real = torch.mean(F.relu(1.0 - logits_real)) + loss_fake = torch.mean(F.relu(1.0 + logits_fake)) + d_loss = 0.5 * (loss_real + loss_fake) + return d_loss + + +def vanilla_d_loss(logits_real, logits_fake): + d_loss = 0.5 * ( + torch.mean(torch.nn.functional.softplus(-logits_real)) + + torch.mean(torch.nn.functional.softplus(logits_fake)) + ) + return d_loss diff --git a/sgm/modules/autoencoding/regularizers/__init__.py b/sgm/modules/autoencoding/regularizers/__init__.py new file mode 100644 index 0000000..ff2b181 --- /dev/null +++ b/sgm/modules/autoencoding/regularizers/__init__.py @@ -0,0 +1,31 @@ +from abc import abstractmethod +from typing import Any, Tuple + +import torch +import torch.nn as nn +import torch.nn.functional as F + +from ....modules.distributions.distributions import \ + DiagonalGaussianDistribution +from .base import AbstractRegularizer + + +class DiagonalGaussianRegularizer(AbstractRegularizer): + def __init__(self, sample: bool = True): + super().__init__() + self.sample = sample + + def get_trainable_parameters(self) -> Any: + yield from () + + def forward(self, z: torch.Tensor) -> Tuple[torch.Tensor, dict]: + log = dict() + posterior = DiagonalGaussianDistribution(z) + if self.sample: + z = posterior.sample() + else: + z = posterior.mode() + kl_loss = posterior.kl() + kl_loss = torch.sum(kl_loss) / kl_loss.shape[0] + log["kl_loss"] = kl_loss + return z, log diff --git a/sgm/modules/autoencoding/regularizers/base.py b/sgm/modules/autoencoding/regularizers/base.py new file mode 100644 index 0000000..fca681b --- /dev/null +++ b/sgm/modules/autoencoding/regularizers/base.py @@ -0,0 +1,40 @@ +from abc import abstractmethod +from typing import Any, Tuple + +import torch +import torch.nn.functional as F +from torch import nn + + +class AbstractRegularizer(nn.Module): + def __init__(self): + super().__init__() + + def forward(self, z: torch.Tensor) -> Tuple[torch.Tensor, dict]: + raise NotImplementedError() + + @abstractmethod + def get_trainable_parameters(self) -> Any: + raise NotImplementedError() + + +class IdentityRegularizer(AbstractRegularizer): + def forward(self, z: torch.Tensor) -> Tuple[torch.Tensor, dict]: + return z, dict() + + def get_trainable_parameters(self) -> Any: + yield from () + + +def measure_perplexity( + predicted_indices: torch.Tensor, num_centroids: int +) -> Tuple[torch.Tensor, torch.Tensor]: + # src: https://github.com/karpathy/deep-vector-quantization/blob/main/model.py + # eval cluster perplexity. when perplexity == num_embeddings then all clusters are used exactly equally + encodings = ( + F.one_hot(predicted_indices, num_centroids).float().reshape(-1, num_centroids) + ) + avg_probs = encodings.mean(0) + perplexity = (-(avg_probs * torch.log(avg_probs + 1e-10)).sum()).exp() + cluster_use = torch.sum(avg_probs > 0) + return perplexity, cluster_use diff --git a/sgm/modules/autoencoding/regularizers/quantize.py b/sgm/modules/autoencoding/regularizers/quantize.py new file mode 100644 index 0000000..86a4dbd --- /dev/null +++ b/sgm/modules/autoencoding/regularizers/quantize.py @@ -0,0 +1,487 @@ +import logging +from abc import abstractmethod +from typing import Dict, Iterator, Literal, Optional, Tuple, Union + +import numpy as np +import torch +import torch.nn as nn +import torch.nn.functional as F +from einops import rearrange +from torch import einsum + +from .base import AbstractRegularizer, measure_perplexity + +logpy = logging.getLogger(__name__) + + +class AbstractQuantizer(AbstractRegularizer): + def __init__(self): + super().__init__() + # Define these in your init + # shape (N,) + self.used: Optional[torch.Tensor] + self.re_embed: int + self.unknown_index: Union[Literal["random"], int] + + def remap_to_used(self, inds: torch.Tensor) -> torch.Tensor: + assert self.used is not None, "You need to define used indices for remap" + ishape = inds.shape + assert len(ishape) > 1 + inds = inds.reshape(ishape[0], -1) + used = self.used.to(inds) + match = (inds[:, :, None] == used[None, None, ...]).long() + new = match.argmax(-1) + unknown = match.sum(2) < 1 + if self.unknown_index == "random": + new[unknown] = torch.randint(0, self.re_embed, size=new[unknown].shape).to( + device=new.device + ) + else: + new[unknown] = self.unknown_index + return new.reshape(ishape) + + def unmap_to_all(self, inds: torch.Tensor) -> torch.Tensor: + assert self.used is not None, "You need to define used indices for remap" + ishape = inds.shape + assert len(ishape) > 1 + inds = inds.reshape(ishape[0], -1) + used = self.used.to(inds) + if self.re_embed > self.used.shape[0]: # extra token + inds[inds >= self.used.shape[0]] = 0 # simply set to zero + back = torch.gather(used[None, :][inds.shape[0] * [0], :], 1, inds) + return back.reshape(ishape) + + @abstractmethod + def get_codebook_entry( + self, indices: torch.Tensor, shape: Optional[Tuple[int, ...]] = None + ) -> torch.Tensor: + raise NotImplementedError() + + def get_trainable_parameters(self) -> Iterator[torch.nn.Parameter]: + yield from self.parameters() + + +class GumbelQuantizer(AbstractQuantizer): + """ + credit to @karpathy: + https://github.com/karpathy/deep-vector-quantization/blob/main/model.py (thanks!) + Gumbel Softmax trick quantizer + Categorical Reparameterization with Gumbel-Softmax, Jang et al. 2016 + https://arxiv.org/abs/1611.01144 + """ + + def __init__( + self, + num_hiddens: int, + embedding_dim: int, + n_embed: int, + straight_through: bool = True, + kl_weight: float = 5e-4, + temp_init: float = 1.0, + remap: Optional[str] = None, + unknown_index: str = "random", + loss_key: str = "loss/vq", + ) -> None: + super().__init__() + + self.loss_key = loss_key + self.embedding_dim = embedding_dim + self.n_embed = n_embed + + self.straight_through = straight_through + self.temperature = temp_init + self.kl_weight = kl_weight + + self.proj = nn.Conv2d(num_hiddens, n_embed, 1) + self.embed = nn.Embedding(n_embed, embedding_dim) + + self.remap = remap + if self.remap is not None: + self.register_buffer("used", torch.tensor(np.load(self.remap))) + self.re_embed = self.used.shape[0] + else: + self.used = None + self.re_embed = n_embed + if unknown_index == "extra": + self.unknown_index = self.re_embed + self.re_embed = self.re_embed + 1 + else: + assert unknown_index == "random" or isinstance( + unknown_index, int + ), "unknown index needs to be 'random', 'extra' or any integer" + self.unknown_index = unknown_index # "random" or "extra" or integer + if self.remap is not None: + logpy.info( + f"Remapping {self.n_embed} indices to {self.re_embed} indices. " + f"Using {self.unknown_index} for unknown indices." + ) + + def forward( + self, z: torch.Tensor, temp: Optional[float] = None, return_logits: bool = False + ) -> Tuple[torch.Tensor, Dict]: + # force hard = True when we are in eval mode, as we must quantize. + # actually, always true seems to work + hard = self.straight_through if self.training else True + temp = self.temperature if temp is None else temp + out_dict = {} + logits = self.proj(z) + if self.remap is not None: + # continue only with used logits + full_zeros = torch.zeros_like(logits) + logits = logits[:, self.used, ...] + + soft_one_hot = F.gumbel_softmax(logits, tau=temp, dim=1, hard=hard) + if self.remap is not None: + # go back to all entries but unused set to zero + full_zeros[:, self.used, ...] = soft_one_hot + soft_one_hot = full_zeros + z_q = einsum("b n h w, n d -> b d h w", soft_one_hot, self.embed.weight) + + # + kl divergence to the prior loss + qy = F.softmax(logits, dim=1) + diff = ( + self.kl_weight + * torch.sum(qy * torch.log(qy * self.n_embed + 1e-10), dim=1).mean() + ) + out_dict[self.loss_key] = diff + + ind = soft_one_hot.argmax(dim=1) + out_dict["indices"] = ind + if self.remap is not None: + ind = self.remap_to_used(ind) + + if return_logits: + out_dict["logits"] = logits + + return z_q, out_dict + + def get_codebook_entry(self, indices, shape): + # TODO: shape not yet optional + b, h, w, c = shape + assert b * h * w == indices.shape[0] + indices = rearrange(indices, "(b h w) -> b h w", b=b, h=h, w=w) + if self.remap is not None: + indices = self.unmap_to_all(indices) + one_hot = ( + F.one_hot(indices, num_classes=self.n_embed).permute(0, 3, 1, 2).float() + ) + z_q = einsum("b n h w, n d -> b d h w", one_hot, self.embed.weight) + return z_q + + +class VectorQuantizer(AbstractQuantizer): + """ + ____________________________________________ + Discretization bottleneck part of the VQ-VAE. + Inputs: + - n_e : number of embeddings + - e_dim : dimension of embedding + - beta : commitment cost used in loss term, + beta * ||z_e(x)-sg[e]||^2 + _____________________________________________ + """ + + def __init__( + self, + n_e: int, + e_dim: int, + beta: float = 0.25, + remap: Optional[str] = None, + unknown_index: str = "random", + sane_index_shape: bool = False, + log_perplexity: bool = False, + embedding_weight_norm: bool = False, + loss_key: str = "loss/vq", + ): + super().__init__() + self.n_e = n_e + self.e_dim = e_dim + self.beta = beta + self.loss_key = loss_key + + if not embedding_weight_norm: + self.embedding = nn.Embedding(self.n_e, self.e_dim) + self.embedding.weight.data.uniform_(-1.0 / self.n_e, 1.0 / self.n_e) + else: + self.embedding = torch.nn.utils.weight_norm( + nn.Embedding(self.n_e, self.e_dim), dim=1 + ) + + self.remap = remap + if self.remap is not None: + self.register_buffer("used", torch.tensor(np.load(self.remap))) + self.re_embed = self.used.shape[0] + else: + self.used = None + self.re_embed = n_e + if unknown_index == "extra": + self.unknown_index = self.re_embed + self.re_embed = self.re_embed + 1 + else: + assert unknown_index == "random" or isinstance( + unknown_index, int + ), "unknown index needs to be 'random', 'extra' or any integer" + self.unknown_index = unknown_index # "random" or "extra" or integer + if self.remap is not None: + logpy.info( + f"Remapping {self.n_e} indices to {self.re_embed} indices. " + f"Using {self.unknown_index} for unknown indices." + ) + + self.sane_index_shape = sane_index_shape + self.log_perplexity = log_perplexity + + def forward( + self, + z: torch.Tensor, + ) -> Tuple[torch.Tensor, Dict]: + do_reshape = z.ndim == 4 + if do_reshape: + # # reshape z -> (batch, height, width, channel) and flatten + z = rearrange(z, "b c h w -> b h w c").contiguous() + + else: + assert z.ndim < 4, "No reshaping strategy for inputs > 4 dimensions defined" + z = z.contiguous() + + z_flattened = z.view(-1, self.e_dim) + # distances from z to embeddings e_j (z - e)^2 = z^2 + e^2 - 2 e * z + + d = ( + torch.sum(z_flattened**2, dim=1, keepdim=True) + + torch.sum(self.embedding.weight**2, dim=1) + - 2 + * torch.einsum( + "bd,dn->bn", z_flattened, rearrange(self.embedding.weight, "n d -> d n") + ) + ) + + min_encoding_indices = torch.argmin(d, dim=1) + z_q = self.embedding(min_encoding_indices).view(z.shape) + loss_dict = {} + if self.log_perplexity: + perplexity, cluster_usage = measure_perplexity( + min_encoding_indices.detach(), self.n_e + ) + loss_dict.update({"perplexity": perplexity, "cluster_usage": cluster_usage}) + + # compute loss for embedding + loss = self.beta * torch.mean((z_q.detach() - z) ** 2) + torch.mean( + (z_q - z.detach()) ** 2 + ) + loss_dict[self.loss_key] = loss + + # preserve gradients + z_q = z + (z_q - z).detach() + + # reshape back to match original input shape + if do_reshape: + z_q = rearrange(z_q, "b h w c -> b c h w").contiguous() + + if self.remap is not None: + min_encoding_indices = min_encoding_indices.reshape( + z.shape[0], -1 + ) # add batch axis + min_encoding_indices = self.remap_to_used(min_encoding_indices) + min_encoding_indices = min_encoding_indices.reshape(-1, 1) # flatten + + if self.sane_index_shape: + if do_reshape: + min_encoding_indices = min_encoding_indices.reshape( + z_q.shape[0], z_q.shape[2], z_q.shape[3] + ) + else: + min_encoding_indices = rearrange( + min_encoding_indices, "(b s) 1 -> b s", b=z_q.shape[0] + ) + + loss_dict["min_encoding_indices"] = min_encoding_indices + + return z_q, loss_dict + + def get_codebook_entry( + self, indices: torch.Tensor, shape: Optional[Tuple[int, ...]] = None + ) -> torch.Tensor: + # shape specifying (batch, height, width, channel) + if self.remap is not None: + assert shape is not None, "Need to give shape for remap" + indices = indices.reshape(shape[0], -1) # add batch axis + indices = self.unmap_to_all(indices) + indices = indices.reshape(-1) # flatten again + + # get quantized latent vectors + z_q = self.embedding(indices) + + if shape is not None: + z_q = z_q.view(shape) + # reshape back to match original input shape + z_q = z_q.permute(0, 3, 1, 2).contiguous() + + return z_q + + +class EmbeddingEMA(nn.Module): + def __init__(self, num_tokens, codebook_dim, decay=0.99, eps=1e-5): + super().__init__() + self.decay = decay + self.eps = eps + weight = torch.randn(num_tokens, codebook_dim) + self.weight = nn.Parameter(weight, requires_grad=False) + self.cluster_size = nn.Parameter(torch.zeros(num_tokens), requires_grad=False) + self.embed_avg = nn.Parameter(weight.clone(), requires_grad=False) + self.update = True + + def forward(self, embed_id): + return F.embedding(embed_id, self.weight) + + def cluster_size_ema_update(self, new_cluster_size): + self.cluster_size.data.mul_(self.decay).add_( + new_cluster_size, alpha=1 - self.decay + ) + + def embed_avg_ema_update(self, new_embed_avg): + self.embed_avg.data.mul_(self.decay).add_(new_embed_avg, alpha=1 - self.decay) + + def weight_update(self, num_tokens): + n = self.cluster_size.sum() + smoothed_cluster_size = ( + (self.cluster_size + self.eps) / (n + num_tokens * self.eps) * n + ) + # normalize embedding average with smoothed cluster size + embed_normalized = self.embed_avg / smoothed_cluster_size.unsqueeze(1) + self.weight.data.copy_(embed_normalized) + + +class EMAVectorQuantizer(AbstractQuantizer): + def __init__( + self, + n_embed: int, + embedding_dim: int, + beta: float, + decay: float = 0.99, + eps: float = 1e-5, + remap: Optional[str] = None, + unknown_index: str = "random", + loss_key: str = "loss/vq", + ): + super().__init__() + self.codebook_dim = embedding_dim + self.num_tokens = n_embed + self.beta = beta + self.loss_key = loss_key + + self.embedding = EmbeddingEMA(self.num_tokens, self.codebook_dim, decay, eps) + + self.remap = remap + if self.remap is not None: + self.register_buffer("used", torch.tensor(np.load(self.remap))) + self.re_embed = self.used.shape[0] + else: + self.used = None + self.re_embed = n_embed + if unknown_index == "extra": + self.unknown_index = self.re_embed + self.re_embed = self.re_embed + 1 + else: + assert unknown_index == "random" or isinstance( + unknown_index, int + ), "unknown index needs to be 'random', 'extra' or any integer" + self.unknown_index = unknown_index # "random" or "extra" or integer + if self.remap is not None: + logpy.info( + f"Remapping {self.n_embed} indices to {self.re_embed} indices. " + f"Using {self.unknown_index} for unknown indices." + ) + + def forward(self, z: torch.Tensor) -> Tuple[torch.Tensor, Dict]: + # reshape z -> (batch, height, width, channel) and flatten + # z, 'b c h w -> b h w c' + z = rearrange(z, "b c h w -> b h w c") + z_flattened = z.reshape(-1, self.codebook_dim) + + # distances from z to embeddings e_j (z - e)^2 = z^2 + e^2 - 2 e * z + d = ( + z_flattened.pow(2).sum(dim=1, keepdim=True) + + self.embedding.weight.pow(2).sum(dim=1) + - 2 * torch.einsum("bd,nd->bn", z_flattened, self.embedding.weight) + ) # 'n d -> d n' + + encoding_indices = torch.argmin(d, dim=1) + + z_q = self.embedding(encoding_indices).view(z.shape) + encodings = F.one_hot(encoding_indices, self.num_tokens).type(z.dtype) + avg_probs = torch.mean(encodings, dim=0) + perplexity = torch.exp(-torch.sum(avg_probs * torch.log(avg_probs + 1e-10))) + + if self.training and self.embedding.update: + # EMA cluster size + encodings_sum = encodings.sum(0) + self.embedding.cluster_size_ema_update(encodings_sum) + # EMA embedding average + embed_sum = encodings.transpose(0, 1) @ z_flattened + self.embedding.embed_avg_ema_update(embed_sum) + # normalize embed_avg and update weight + self.embedding.weight_update(self.num_tokens) + + # compute loss for embedding + loss = self.beta * F.mse_loss(z_q.detach(), z) + + # preserve gradients + z_q = z + (z_q - z).detach() + + # reshape back to match original input shape + # z_q, 'b h w c -> b c h w' + z_q = rearrange(z_q, "b h w c -> b c h w") + + out_dict = { + self.loss_key: loss, + "encodings": encodings, + "encoding_indices": encoding_indices, + "perplexity": perplexity, + } + + return z_q, out_dict + + +class VectorQuantizerWithInputProjection(VectorQuantizer): + def __init__( + self, + input_dim: int, + n_codes: int, + codebook_dim: int, + beta: float = 1.0, + output_dim: Optional[int] = None, + **kwargs, + ): + super().__init__(n_codes, codebook_dim, beta, **kwargs) + self.proj_in = nn.Linear(input_dim, codebook_dim) + self.output_dim = output_dim + if output_dim is not None: + self.proj_out = nn.Linear(codebook_dim, output_dim) + else: + self.proj_out = nn.Identity() + + def forward(self, z: torch.Tensor) -> Tuple[torch.Tensor, Dict]: + rearr = False + in_shape = z.shape + + if z.ndim > 3: + rearr = self.output_dim is not None + z = rearrange(z, "b c ... -> b (...) c") + z = self.proj_in(z) + z_q, loss_dict = super().forward(z) + + z_q = self.proj_out(z_q) + if rearr: + if len(in_shape) == 4: + z_q = rearrange(z_q, "b (h w) c -> b c h w ", w=in_shape[-1]) + elif len(in_shape) == 5: + z_q = rearrange( + z_q, "b (t h w) c -> b c t h w ", w=in_shape[-1], h=in_shape[-2] + ) + else: + raise NotImplementedError( + f"rearranging not available for {len(in_shape)}-dimensional input." + ) + + return z_q, loss_dict diff --git a/sgm/modules/autoencoding/temporal_ae.py b/sgm/modules/autoencoding/temporal_ae.py new file mode 100644 index 0000000..4a17a91 --- /dev/null +++ b/sgm/modules/autoencoding/temporal_ae.py @@ -0,0 +1,347 @@ +from typing import Callable, Iterable, Union + +import torch +from einops import rearrange, repeat + +from sgm.modules.diffusionmodules.model import (XFORMERS_IS_AVAILABLE, + AttnBlock, Decoder, + MemoryEfficientAttnBlock, + ResnetBlock) +from sgm.modules.diffusionmodules.openaimodel import (ResBlock, + timestep_embedding) +from sgm.modules.video_attention import VideoTransformerBlock +from sgm.util import partialclass + + +class VideoResBlock(ResnetBlock): + def __init__( + self, + out_channels, + *args, + dropout=0.0, + video_kernel_size=3, + alpha=0.0, + merge_strategy="learned", + **kwargs, + ): + super().__init__(out_channels=out_channels, dropout=dropout, *args, **kwargs) + if video_kernel_size is None: + video_kernel_size = [3, 1, 1] + self.time_stack = ResBlock( + channels=out_channels, + emb_channels=0, + dropout=dropout, + dims=3, + use_scale_shift_norm=False, + use_conv=False, + up=False, + down=False, + kernel_size=video_kernel_size, + use_checkpoint=False, + skip_t_emb=True, + ) + + self.merge_strategy = merge_strategy + if self.merge_strategy == "fixed": + self.register_buffer("mix_factor", torch.Tensor([alpha])) + elif self.merge_strategy == "learned": + self.register_parameter( + "mix_factor", torch.nn.Parameter(torch.Tensor([alpha])) + ) + else: + raise ValueError(f"unknown merge strategy {self.merge_strategy}") + + def get_alpha(self, bs): + if self.merge_strategy == "fixed": + return self.mix_factor + elif self.merge_strategy == "learned": + return torch.sigmoid(self.mix_factor) + else: + raise NotImplementedError() + + def forward(self, x, temb, skip_video=False, timesteps=None): + if timesteps is None: + timesteps = self.timesteps + + b, c, h, w = x.shape + + x = super().forward(x, temb) + + if not skip_video: + x_mix = rearrange(x, "(b t) c h w -> b c t h w", t=timesteps) + + x = rearrange(x, "(b t) c h w -> b c t h w", t=timesteps) + + x = self.time_stack(x, temb) + + alpha = self.get_alpha(bs=b // timesteps) + x = alpha * x + (1.0 - alpha) * x_mix + + x = rearrange(x, "b c t h w -> (b t) c h w") + return x + + +class AE3DConv(torch.nn.Conv2d): + def __init__(self, in_channels, out_channels, video_kernel_size=3, *args, **kwargs): + super().__init__(in_channels, out_channels, *args, **kwargs) + if isinstance(video_kernel_size, Iterable): + padding = [int(k // 2) for k in video_kernel_size] + else: + padding = int(video_kernel_size // 2) + + self.time_mix_conv = torch.nn.Conv3d( + in_channels=out_channels, + out_channels=out_channels, + kernel_size=video_kernel_size, + padding=padding, + ) + + def forward(self, input, timesteps, skip_video=False): + x = super().forward(input) + if skip_video: + return x + x = rearrange(x, "(b t) c h w -> b c t h w", t=timesteps) + x = self.time_mix_conv(x) + return rearrange(x, "b c t h w -> (b t) c h w") + + +class VideoBlock(AttnBlock): + def __init__( + self, in_channels: int, alpha: float = 0, merge_strategy: str = "learned" + ): + super().__init__(in_channels) + # no context, single headed, as in base class + self.time_mix_block = VideoTransformerBlock( + dim=in_channels, + n_heads=1, + d_head=in_channels, + checkpoint=False, + ff_in=True, + attn_mode="softmax", + ) + + time_embed_dim = self.in_channels * 4 + self.video_time_embed = torch.nn.Sequential( + torch.nn.Linear(self.in_channels, time_embed_dim), + torch.nn.SiLU(), + torch.nn.Linear(time_embed_dim, self.in_channels), + ) + + self.merge_strategy = merge_strategy + if self.merge_strategy == "fixed": + self.register_buffer("mix_factor", torch.Tensor([alpha])) + elif self.merge_strategy == "learned": + self.register_parameter( + "mix_factor", torch.nn.Parameter(torch.Tensor([alpha])) + ) + else: + raise ValueError(f"unknown merge strategy {self.merge_strategy}") + + def forward(self, x, timesteps, skip_video=False): + if skip_video: + return super().forward(x) + + x_in = x + x = self.attention(x) + h, w = x.shape[2:] + x = rearrange(x, "b c h w -> b (h w) c") + + x_mix = x + num_frames = torch.arange(timesteps, device=x.device) + num_frames = repeat(num_frames, "t -> b t", b=x.shape[0] // timesteps) + num_frames = rearrange(num_frames, "b t -> (b t)") + t_emb = timestep_embedding(num_frames, self.in_channels, repeat_only=False) + emb = self.video_time_embed(t_emb) # b, n_channels + emb = emb[:, None, :] + x_mix = x_mix + emb + + alpha = self.get_alpha() + x_mix = self.time_mix_block(x_mix, timesteps=timesteps) + x = alpha * x + (1.0 - alpha) * x_mix # alpha merge + + x = rearrange(x, "b (h w) c -> b c h w", h=h, w=w) + x = self.proj_out(x) + + return x_in + x + + def get_alpha( + self, + ): + if self.merge_strategy == "fixed": + return self.mix_factor + elif self.merge_strategy == "learned": + return torch.sigmoid(self.mix_factor) + else: + raise NotImplementedError(f"unknown merge strategy {self.merge_strategy}") + + +class MemoryEfficientVideoBlock(MemoryEfficientAttnBlock): + def __init__( + self, in_channels: int, alpha: float = 0, merge_strategy: str = "learned" + ): + super().__init__(in_channels) + # no context, single headed, as in base class + self.time_mix_block = VideoTransformerBlock( + dim=in_channels, + n_heads=1, + d_head=in_channels, + checkpoint=False, + ff_in=True, + attn_mode="softmax-xformers", + ) + + time_embed_dim = self.in_channels * 4 + self.video_time_embed = torch.nn.Sequential( + torch.nn.Linear(self.in_channels, time_embed_dim), + torch.nn.SiLU(), + torch.nn.Linear(time_embed_dim, self.in_channels), + ) + + self.merge_strategy = merge_strategy + if self.merge_strategy == "fixed": + self.register_buffer("mix_factor", torch.Tensor([alpha])) + elif self.merge_strategy == "learned": + self.register_parameter( + "mix_factor", torch.nn.Parameter(torch.Tensor([alpha])) + ) + else: + raise ValueError(f"unknown merge strategy {self.merge_strategy}") + + def forward(self, x, timesteps, skip_time_block=False): + if skip_time_block: + return super().forward(x) + + x_in = x + x = self.attention(x) + h, w = x.shape[2:] + x = rearrange(x, "b c h w -> b (h w) c") + + x_mix = x + num_frames = torch.arange(timesteps, device=x.device) + num_frames = repeat(num_frames, "t -> b t", b=x.shape[0] // timesteps) + num_frames = rearrange(num_frames, "b t -> (b t)") + t_emb = timestep_embedding(num_frames, self.in_channels, repeat_only=False) + emb = self.video_time_embed(t_emb) # b, n_channels + emb = emb[:, None, :] + x_mix = x_mix + emb + + alpha = self.get_alpha() + x_mix = self.time_mix_block(x_mix, timesteps=timesteps) + x = alpha * x + (1.0 - alpha) * x_mix # alpha merge + + x = rearrange(x, "b (h w) c -> b c h w", h=h, w=w) + x = self.proj_out(x) + + return x_in + x + + def get_alpha( + self, + ): + if self.merge_strategy == "fixed": + return self.mix_factor + elif self.merge_strategy == "learned": + return torch.sigmoid(self.mix_factor) + else: + raise NotImplementedError(f"unknown merge strategy {self.merge_strategy}") + + +def make_time_attn( + in_channels, + attn_type="vanilla", + attn_kwargs=None, + alpha: float = 0, + merge_strategy: str = "learned", +): + assert attn_type in [ + "vanilla", + "vanilla-xformers", + ], f"attn_type {attn_type} not supported for spatio-temporal attention" + print( + f"making spatial and temporal attention of type '{attn_type}' with {in_channels} in_channels" + ) + if not XFORMERS_IS_AVAILABLE and attn_type == "vanilla-xformers": + print( + f"Attention mode '{attn_type}' is not available. Falling back to vanilla attention. " + f"This is not a problem in Pytorch >= 2.0. FYI, you are running with PyTorch version {torch.__version__}" + ) + attn_type = "vanilla" + + if attn_type == "vanilla": + assert attn_kwargs is None + return partialclass( + VideoBlock, in_channels, alpha=alpha, merge_strategy=merge_strategy + ) + elif attn_type == "vanilla-xformers": + print(f"building MemoryEfficientAttnBlock with {in_channels} in_channels...") + return partialclass( + MemoryEfficientVideoBlock, + in_channels, + alpha=alpha, + merge_strategy=merge_strategy, + ) + else: + return NotImplementedError() + + +class Conv2DWrapper(torch.nn.Conv2d): + def forward(self, input: torch.Tensor, **kwargs) -> torch.Tensor: + return super().forward(input) + + +class VideoDecoder(Decoder): + available_time_modes = ["all", "conv-only", "attn-only"] + + def __init__( + self, + *args, + video_kernel_size: Union[int, list] = 3, + alpha: float = 0.0, + merge_strategy: str = "learned", + time_mode: str = "conv-only", + **kwargs, + ): + self.video_kernel_size = video_kernel_size + self.alpha = alpha + self.merge_strategy = merge_strategy + self.time_mode = time_mode + assert ( + self.time_mode in self.available_time_modes + ), f"time_mode parameter has to be in {self.available_time_modes}" + super().__init__(*args, **kwargs) + + def get_last_layer(self, skip_time_mix=False, **kwargs): + if self.time_mode == "attn-only": + raise NotImplementedError("TODO") + else: + return ( + self.conv_out.time_mix_conv.weight + if not skip_time_mix + else self.conv_out.weight + ) + + def _make_attn(self) -> Callable: + if self.time_mode not in ["conv-only", "only-last-conv"]: + return partialclass( + make_time_attn, + alpha=self.alpha, + merge_strategy=self.merge_strategy, + ) + else: + return super()._make_attn() + + def _make_conv(self) -> Callable: + if self.time_mode != "attn-only": + return partialclass(AE3DConv, video_kernel_size=self.video_kernel_size) + else: + return Conv2DWrapper + + def _make_resblock(self) -> Callable: + if self.time_mode not in ["attn-only", "only-last-conv"]: + return partialclass( + VideoResBlock, + video_kernel_size=self.video_kernel_size, + alpha=self.alpha, + merge_strategy=self.merge_strategy, + ) + else: + return super()._make_resblock() diff --git a/sgm/modules/diffusionmodules/__init__.py b/sgm/modules/diffusionmodules/__init__.py new file mode 100644 index 0000000..e69de29 diff --git a/sgm/modules/diffusionmodules/denoiser.py b/sgm/modules/diffusionmodules/denoiser.py new file mode 100644 index 0000000..f39404e --- /dev/null +++ b/sgm/modules/diffusionmodules/denoiser.py @@ -0,0 +1,80 @@ +from typing import Dict, Union + +import torch +import torch.nn as nn + +from ...util import append_dims, instantiate_from_config +from .denoiser_scaling import DenoiserScaling +from .discretizer import Discretization + + +class Denoiser(nn.Module): + def __init__(self, scaling_config: Dict): + super().__init__() + + self.scaling: DenoiserScaling = instantiate_from_config(scaling_config) + + def possibly_quantize_sigma(self, sigma: torch.Tensor) -> torch.Tensor: + return sigma + + def possibly_quantize_c_noise(self, c_noise: torch.Tensor) -> torch.Tensor: + return c_noise + + def forward( + self, + network: nn.Module, + input: torch.Tensor, + sigma: torch.Tensor, + cond: Dict, + **additional_model_inputs, + ) -> torch.Tensor: + sigma = self.possibly_quantize_sigma(sigma) + sigma_shape = sigma.shape + sigma = append_dims(sigma, input.ndim) + c_skip, c_out, c_in, c_noise = self.scaling(sigma) + c_noise = self.possibly_quantize_c_noise(c_noise.reshape(sigma_shape)) + + if "replace" in cond: + x, mask = cond["replace"].split((input.shape[-3], 1), dim=-3) # replace with clean latent replace with clean latent + input = input * (1 - mask) + x * mask + + return ( + network(input * c_in, c_noise, cond, **additional_model_inputs) * c_out + + input * c_skip + ) + + +class DiscreteDenoiser(Denoiser): + def __init__( + self, + scaling_config: Dict, + num_idx: int, + discretization_config: Dict, + do_append_zero: bool = False, + quantize_c_noise: bool = True, + flip: bool = True, + ): + super().__init__(scaling_config) + self.discretization: Discretization = instantiate_from_config( + discretization_config + ) + sigmas = self.discretization(num_idx, do_append_zero=do_append_zero, flip=flip) + self.register_buffer("sigmas", sigmas) + self.quantize_c_noise = quantize_c_noise + self.num_idx = num_idx + + def sigma_to_idx(self, sigma: torch.Tensor) -> torch.Tensor: + dists = sigma - self.sigmas[:, None] + return dists.abs().argmin(dim=0).view(sigma.shape) + + def idx_to_sigma(self, idx: Union[torch.Tensor, int]) -> torch.Tensor: + return self.sigmas[idx] + + def possibly_quantize_sigma(self, sigma: torch.Tensor) -> torch.Tensor: + return self.idx_to_sigma(self.sigma_to_idx(sigma)) + + def possibly_quantize_c_noise(self, c_noise: torch.Tensor) -> torch.Tensor: + if self.quantize_c_noise: + return self.sigma_to_idx(c_noise) + else: + return c_noise diff --git a/sgm/modules/diffusionmodules/denoiser_scaling.py b/sgm/modules/diffusionmodules/denoiser_scaling.py new file mode 100644 index 0000000..f4e287b --- /dev/null +++ b/sgm/modules/diffusionmodules/denoiser_scaling.py @@ -0,0 +1,59 @@ +from abc import ABC, abstractmethod +from typing import Tuple + +import torch + + +class DenoiserScaling(ABC): + @abstractmethod + def __call__( + self, sigma: torch.Tensor + ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]: + pass + + +class EDMScaling: + def __init__(self, sigma_data: float = 0.5): + self.sigma_data = sigma_data + + def __call__( + self, sigma: torch.Tensor + ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]: + c_skip = self.sigma_data**2 / (sigma**2 + self.sigma_data**2) + c_out = sigma * self.sigma_data / (sigma**2 + self.sigma_data**2) ** 0.5 + c_in = 1 / (sigma**2 + self.sigma_data**2) ** 0.5 + c_noise = 0.25 * sigma.log() + return c_skip, c_out, c_in, c_noise + + +class EpsScaling: + def __call__( + self, sigma: torch.Tensor + ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]: + c_skip = torch.ones_like(sigma, device=sigma.device) + c_out = -sigma + c_in = 1 / (sigma**2 + 1.0) ** 0.5 + c_noise = sigma.clone() + return c_skip, c_out, c_in, c_noise + + +class VScaling: + def __call__( + self, sigma: torch.Tensor + ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]: + c_skip = 1.0 / (sigma**2 + 1.0) + c_out = -sigma / (sigma**2 + 1.0) ** 0.5 + c_in = 1.0 / (sigma**2 + 1.0) ** 0.5 + c_noise = sigma.clone() + return c_skip, c_out, c_in, c_noise + + +class VScalingWithEDMcNoise(DenoiserScaling): + def __call__( + self, sigma: torch.Tensor + ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]: + c_skip = 1.0 / (sigma**2 + 1.0) + c_out = -sigma / (sigma**2 + 1.0) ** 0.5 + c_in = 1.0 / (sigma**2 + 1.0) ** 0.5 + c_noise = 0.25 * sigma.log() + return c_skip, c_out, c_in, c_noise diff --git a/sgm/modules/diffusionmodules/denoiser_weighting.py b/sgm/modules/diffusionmodules/denoiser_weighting.py new file mode 100644 index 0000000..b8b03ca --- /dev/null +++ b/sgm/modules/diffusionmodules/denoiser_weighting.py @@ -0,0 +1,24 @@ +import torch + + +class UnitWeighting: + def __call__(self, sigma): + return torch.ones_like(sigma, device=sigma.device) + + +class EDMWeighting: + def __init__(self, sigma_data=0.5): + self.sigma_data = sigma_data + + def __call__(self, sigma): + return (sigma**2 + self.sigma_data**2) / (sigma * self.sigma_data) ** 2 + + +class VWeighting(EDMWeighting): + def __init__(self): + super().__init__(sigma_data=1.0) + + +class EpsWeighting: + def __call__(self, sigma): + return sigma**-2.0 diff --git a/sgm/modules/diffusionmodules/discretizer.py b/sgm/modules/diffusionmodules/discretizer.py new file mode 100644 index 0000000..02add60 --- /dev/null +++ b/sgm/modules/diffusionmodules/discretizer.py @@ -0,0 +1,69 @@ +from abc import abstractmethod +from functools import partial + +import numpy as np +import torch + +from ...modules.diffusionmodules.util import make_beta_schedule +from ...util import append_zero + + +def generate_roughly_equally_spaced_steps( + num_substeps: int, max_step: int +) -> np.ndarray: + return np.linspace(max_step - 1, 0, num_substeps, endpoint=False).astype(int)[::-1] + + +class Discretization: + def __call__(self, n, do_append_zero=True, device="cpu", flip=False): + sigmas = self.get_sigmas(n, device=device) + sigmas = append_zero(sigmas) if do_append_zero else sigmas + return sigmas if not flip else torch.flip(sigmas, (0,)) + + @abstractmethod + def get_sigmas(self, n, device): + pass + + +class EDMDiscretization(Discretization): + def __init__(self, sigma_min=0.002, sigma_max=80.0, rho=7.0): + self.sigma_min = sigma_min + self.sigma_max = sigma_max + self.rho = rho + + def get_sigmas(self, n, device="cpu"): + ramp = torch.linspace(0, 1, n, device=device) + min_inv_rho = self.sigma_min ** (1 / self.rho) + max_inv_rho = self.sigma_max ** (1 / self.rho) + sigmas = (max_inv_rho + ramp * (min_inv_rho - max_inv_rho)) ** self.rho + return sigmas + + +class LegacyDDPMDiscretization(Discretization): + def __init__( + self, + linear_start=0.00085, + linear_end=0.0120, + num_timesteps=1000, + ): + super().__init__() + self.num_timesteps = num_timesteps + betas = make_beta_schedule( + "linear", num_timesteps, linear_start=linear_start, linear_end=linear_end + ) + alphas = 1.0 - betas + self.alphas_cumprod = np.cumprod(alphas, axis=0) + self.to_torch = partial(torch.tensor, dtype=torch.float32) + + def get_sigmas(self, n, device="cpu"): + if n < self.num_timesteps: + timesteps = generate_roughly_equally_spaced_steps(n, self.num_timesteps) + alphas_cumprod = self.alphas_cumprod[timesteps] + elif n == self.num_timesteps: + alphas_cumprod = self.alphas_cumprod + else: + raise ValueError + + to_torch = partial(torch.tensor, dtype=torch.float32, device=device) + sigmas = to_torch((1 - alphas_cumprod) / alphas_cumprod) ** 0.5 + return torch.flip(sigmas, (0,)) diff --git a/sgm/modules/diffusionmodules/guiders.py b/sgm/modules/diffusionmodules/guiders.py new file mode 100644 index 0000000..3589344 --- /dev/null +++ b/sgm/modules/diffusionmodules/guiders.py @@ -0,0 +1,175 @@ +import logging +from abc import ABC, abstractmethod +from typing import Dict, List, Literal, Optional, Tuple, Union + +import torch +from einops import rearrange, repeat + +from ...util import append_dims, default + +logpy = logging.getLogger(__name__) + + +class Guider(ABC): + @abstractmethod + def __call__(self, x: torch.Tensor, sigma: float) -> torch.Tensor: + pass + + def prepare_inputs( + self, x: torch.Tensor, s: float, c: Dict, uc: Dict + ) -> Tuple[torch.Tensor, float, Dict]: + pass + + +class VanillaCFG(Guider): + def __init__(self, scale: float): + self.scale = scale + + def __call__(self, x: torch.Tensor, sigma: torch.Tensor) -> torch.Tensor: + x_u, x_c = x.chunk(2) + x_pred = x_u + self.scale * (x_c - x_u) + return x_pred + + def prepare_inputs(self, x, s, c, uc): + c_out = dict() + + for k in c: + if k in ["vector", "crossattn", "concat", "mask", "plucker"]: + c_out[k] = torch.cat((uc[k], c[k]), 0) + else: + assert c[k] == uc[k] + c_out[k] = c[k] + return torch.cat([x] * 2), torch.cat([s] * 2), c_out + + +class IdentityGuider(Guider): + def __call__(self, x: torch.Tensor, sigma: float) -> torch.Tensor: + return x + + def prepare_inputs( + self, x: torch.Tensor, s: float, c: Dict, uc: Dict + ) -> Tuple[torch.Tensor, float, Dict]: + c_out = dict() + + for k in c: + c_out[k] = c[k] + + return x, s, c_out + + +class LinearPredictionGuider(Guider): + def __init__( + self, + max_scale: float, + num_frames: int, + min_scale: float = 1.0, + additional_cond_keys: Optional[Union[List[str], str]] = None, + ): + self.min_scale = min_scale + self.max_scale = max_scale + self.num_frames = num_frames + self.scale = torch.linspace(min_scale, max_scale, num_frames).unsqueeze(0) + + additional_cond_keys = default(additional_cond_keys, []) + if isinstance(additional_cond_keys, str): + additional_cond_keys = [additional_cond_keys] + self.additional_cond_keys = additional_cond_keys + + def __call__(self, x: torch.Tensor, sigma: torch.Tensor) -> torch.Tensor: + x_u, x_c = x.chunk(2) + + x_u = rearrange(x_u, "(b t) ... -> b t ...", t=self.num_frames) + x_c = rearrange(x_c, "(b t) ... -> b t ...", t=self.num_frames) + scale = repeat(self.scale, "1 t -> b t", b=x_u.shape[0]) + scale = append_dims(scale, x_u.ndim).to(x_u.device) + + return rearrange(x_u + scale * (x_c - x_u), "b t ... -> (b t) ...") + + def prepare_inputs( + self, x: torch.Tensor, s: torch.Tensor, c: dict, uc: dict + ) -> Tuple[torch.Tensor, torch.Tensor, dict]: + c_out = dict() + + for k in c: + if k in ["vector", "crossattn", "concat"] + self.additional_cond_keys: + c_out[k] = torch.cat((uc[k], c[k]), 0) + else: + # assert c[k] == uc[k] + c_out[k] = c[k] + return torch.cat([x] * 2), torch.cat([s] * 2), c_out + + +class TrianglePredictionGuider(LinearPredictionGuider): + def __init__( + self, + max_scale: float, + num_frames: int, + min_scale: float = 1.0, + period: Union[float, List[float]] = 1.0, + period_fusing: Literal["mean", "multiply", "max"] = "max", + additional_cond_keys: Optional[Union[List[str], str]] = None, + ): + super().__init__(max_scale, num_frames, min_scale, additional_cond_keys) + values = torch.linspace(0, 1, num_frames) + # Constructs a triangle wave + if isinstance(period, float): + period = [period] + + scales = [] + for p in period: + scales.append(self.triangle_wave(values, p)) + + if period_fusing == "mean": + scale = sum(scales) / len(period) + elif period_fusing == "multiply": + scale = torch.prod(torch.stack(scales), dim=0) + elif period_fusing == "max": + scale = torch.max(torch.stack(scales), dim=0).values + self.scale = (scale * (max_scale - min_scale) + min_scale).unsqueeze(0) + + def triangle_wave(self, values: torch.Tensor, period) -> torch.Tensor: + return 2 * (values / period - torch.floor(values / period + 0.5)).abs() + + +class TrapezoidPredictionGuider(LinearPredictionGuider): + def __init__( + self, + max_scale: float, + num_frames: int, + min_scale: float = 1.0, + edge_perc: float = 0.1, + additional_cond_keys: Optional[Union[List[str], str]] = None, + ): + super().__init__(max_scale, num_frames, min_scale, additional_cond_keys) + + rise_steps = torch.linspace(min_scale, max_scale, int(num_frames * edge_perc)) + fall_steps = torch.flip(rise_steps, [0]) + self.scale = torch.cat( + [ + rise_steps, + torch.ones(num_frames - 2 * int(num_frames * edge_perc)), + fall_steps, + ] + ).unsqueeze(0) + + +class SpatiotemporalPredictionGuider(LinearPredictionGuider): + def __init__( + self, + max_scale: float, + num_frames: int, + num_views: int = 1, + min_scale: float = 1.0, + additional_cond_keys: Optional[Union[List[str], str]] = None, + ): + super().__init__(max_scale, num_frames, min_scale, additional_cond_keys) + V = num_views + T = num_frames // V + scale = torch.zeros(num_frames).view(T, V) + scale += torch.linspace(0, 1, T)[:,None] * 0.5 + scale += self.triangle_wave(torch.linspace(0, 1, V))[None,:] * 0.5 + scale = scale.flatten() + self.scale = (scale * (max_scale - min_scale) + min_scale).unsqueeze(0) + + def triangle_wave(self, values: torch.Tensor, period=1) -> torch.Tensor: + return 2 * (values / period - torch.floor(values / period + 0.5)).abs() \ No newline at end of file diff --git a/sgm/modules/diffusionmodules/loss.py b/sgm/modules/diffusionmodules/loss.py new file mode 100644 index 0000000..957839f --- /dev/null +++ b/sgm/modules/diffusionmodules/loss.py @@ -0,0 +1,121 @@ +from typing import Dict, List, Optional, Tuple, Union + +import torch +import torch.nn as nn + +from ...modules.autoencoding.lpips.loss.lpips import LPIPS +from ...modules.encoders.modules import GeneralConditioner +from ...util import append_dims, instantiate_from_config +from .denoiser import Denoiser + + +class StandardDiffusionLoss(nn.Module): + def __init__( + self, + sigma_sampler_config: dict, + loss_weighting_config: dict, + loss_type: str = "l2", + offset_noise_level: float = 0.0, + batch2model_keys: Optional[Union[str, List[str]]] = None, + ): + super().__init__() + + assert loss_type in ["l2", "l1", "lpips"] + + self.sigma_sampler = instantiate_from_config(sigma_sampler_config) + self.loss_weighting = instantiate_from_config(loss_weighting_config) + + self.loss_type = loss_type + self.offset_noise_level = offset_noise_level + + if loss_type == "lpips": + self.lpips = LPIPS().eval() + + if not batch2model_keys: + batch2model_keys = [] + + if isinstance(batch2model_keys, str): + batch2model_keys = [batch2model_keys] + + self.batch2model_keys = set(batch2model_keys) + + def get_noised_input( + self, sigmas_bc: torch.Tensor, noise: torch.Tensor, input: torch.Tensor + ) -> torch.Tensor: + noised_input = input + noise * sigmas_bc + return noised_input + + def forward( + self, + network: nn.Module, + denoiser: Denoiser, + conditioner: GeneralConditioner, + input: torch.Tensor, + batch: Dict, + ) -> torch.Tensor: + cond = conditioner(batch) + return self._forward(network, denoiser, cond, input, batch) + + def _forward( + self, + network: nn.Module, + denoiser: Denoiser, + cond: Dict, + input: torch.Tensor, + batch: Dict, + ) -> Tuple[torch.Tensor, Dict]: + additional_model_inputs = { + key: batch[key] for key in self.batch2model_keys.intersection(batch) + } + sigmas = self.sigma_sampler(input.shape[0]).to(input) + + noise = torch.randn_like(input) + if self.offset_noise_level > 0.0: + offset_shape = ( + (input.shape[0], 1, input.shape[2]) + if self.n_frames is not None + else (input.shape[0], input.shape[1]) + ) + noise = noise + self.offset_noise_level * append_dims( + torch.randn(offset_shape, device=input.device), + input.ndim, + ) + sigmas_bc = append_dims(sigmas, input.ndim) + noised_input = self.get_noised_input(sigmas_bc, noise, input) + + model_output = denoiser( + network, noised_input, sigmas, cond, **additional_model_inputs + ) + w = append_dims(self.loss_weighting(sigmas, cond["mask"]), input.ndim) + return self.get_loss(model_output, input, w) + + def get_loss(self, model_output, target, w): + if self.loss_type == "l2": + return torch.mean( + (w * (model_output - target) ** 2).reshape(target.shape[0], -1), 1 + ) + elif self.loss_type == "l1": + return torch.mean( + (w * (model_output - target).abs()).reshape(target.shape[0], -1), 1 + ) + elif self.loss_type == "lpips": + loss = self.lpips(model_output, target).reshape(-1) + return loss + else: + raise NotImplementedError(f"Unknown loss type {self.loss_type}") + +def interpolate_weights_batch(bools: torch.Tensor, max_weight=5.0) -> torch.Tensor: + B, N = bools.shape + indices = torch.arange(N, device=bools.device).unsqueeze(0).expand(B, N) + weights = torch.full((B, N), max_weight, dtype=torch.float, device=bools.device) + + for b in range(B): + true_idx = indices[b][bools[b]] + if len(true_idx) > 0: + dists = torch.stack([torch.abs(indices[b] - t) for t in true_idx]).min(dim=0).values + dists[bools[b]] = 0 + weights[b] = dists / dists.max() * max_weight + else: + weights[b] = max_weight + + return weights \ No newline at end of file diff --git a/sgm/modules/diffusionmodules/loss_weighting.py b/sgm/modules/diffusionmodules/loss_weighting.py new file mode 100644 index 0000000..f26dc17 --- /dev/null +++ b/sgm/modules/diffusionmodules/loss_weighting.py @@ -0,0 +1,51 @@ +from abc import ABC, abstractmethod + +import torch + + +class DiffusionLossWeighting(ABC): + @abstractmethod + def __call__(self, sigma: torch.Tensor) -> torch.Tensor: + pass + + +class UnitWeighting(DiffusionLossWeighting): + def __call__(self, sigma: torch.Tensor) -> torch.Tensor: + return torch.ones_like(sigma, device=sigma.device) + + +class EDMWeighting(DiffusionLossWeighting): + def __init__(self, sigma_data: float = 0.5): + self.sigma_data = sigma_data + + def __call__(self, sigma: torch.Tensor) -> torch.Tensor: + return (sigma**2 + self.sigma_data**2) / (sigma * self.sigma_data) ** 2 + + +class VWeighting(EDMWeighting): + def __init__(self): + super().__init__(sigma_data=1.0) + + +class EpsWeighting(DiffusionLossWeighting): + def __call__(self, sigma: torch.Tensor) -> torch.Tensor: + return sigma**-2.0 + +class SevaWeighting(DiffusionLossWeighting): + def __call__(self, sigma: torch.Tensor, mask, max_weight=5.0) -> torch.Tensor: + bools = mask.to(torch.bool) + batch_size, N = bools.shape + indices = torch.arange(N, device=bools.device).unsqueeze(0).expand(batch_size, N) + weights = torch.full((batch_size, N), max_weight, dtype=torch.float, device=bools.device) + + for b in range(batch_size): + true_idx = indices[b][bools[b]] + if len(true_idx) > 0: + dists = torch.stack([torch.abs(indices[b] - t) for t in true_idx]).min(dim=0).values + dists[bools[b]] = 0 + weights[b] = dists / dists.max() * max_weight + else: + weights[b] = max_weight + + return weights + diff --git a/sgm/modules/diffusionmodules/model.py b/sgm/modules/diffusionmodules/model.py new file mode 100644 index 0000000..4cf9d92 --- /dev/null +++ b/sgm/modules/diffusionmodules/model.py @@ -0,0 +1,748 @@ +# pytorch_diffusion + derived encoder decoder +import logging +import math +from typing import Any, Callable, Optional + +import numpy as np +import torch +import torch.nn as nn +from einops import rearrange +from packaging import version + +logpy = logging.getLogger(__name__) + +try: + import xformers + import xformers.ops + + XFORMERS_IS_AVAILABLE = True +except: + XFORMERS_IS_AVAILABLE = False + logpy.warning("no module 'xformers'. Processing without...") + +from ...modules.attention import LinearAttention, MemoryEfficientCrossAttention + + +def get_timestep_embedding(timesteps, embedding_dim): + """ + This matches the implementation in Denoising Diffusion Probabilistic Models: + From Fairseq. + Build sinusoidal embeddings. + This matches the implementation in tensor2tensor, but differs slightly + from the description in Section 3.5 of "Attention Is All You Need". + """ + assert len(timesteps.shape) == 1 + + half_dim = embedding_dim // 2 + emb = math.log(10000) / (half_dim - 1) + emb = torch.exp(torch.arange(half_dim, dtype=torch.float32) * -emb) + emb = emb.to(device=timesteps.device) + emb = timesteps.float()[:, None] * emb[None, :] + emb = torch.cat([torch.sin(emb), torch.cos(emb)], dim=1) + if embedding_dim % 2 == 1: # zero pad + emb = torch.nn.functional.pad(emb, (0, 1, 0, 0)) + return emb + + +def nonlinearity(x): + # swish + return x * torch.sigmoid(x) + + +def Normalize(in_channels, num_groups=32): + return torch.nn.GroupNorm( + num_groups=num_groups, num_channels=in_channels, eps=1e-6, affine=True + ) + + +class Upsample(nn.Module): + def __init__(self, in_channels, with_conv): + super().__init__() + self.with_conv = with_conv + if self.with_conv: + self.conv = torch.nn.Conv2d( + in_channels, in_channels, kernel_size=3, stride=1, padding=1 + ) + + def forward(self, x): + x = torch.nn.functional.interpolate(x, scale_factor=2.0, mode="nearest") + if self.with_conv: + x = self.conv(x) + return x + + +class Downsample(nn.Module): + def __init__(self, in_channels, with_conv): + super().__init__() + self.with_conv = with_conv + if self.with_conv: + # no asymmetric padding in torch conv, must do it ourselves + self.conv = torch.nn.Conv2d( + in_channels, in_channels, kernel_size=3, stride=2, padding=0 + ) + + def forward(self, x): + if self.with_conv: + pad = (0, 1, 0, 1) + x = torch.nn.functional.pad(x, pad, mode="constant", value=0) + x = self.conv(x) + else: + x = torch.nn.functional.avg_pool2d(x, kernel_size=2, stride=2) + return x + + +class ResnetBlock(nn.Module): + def __init__( + self, + *, + in_channels, + out_channels=None, + conv_shortcut=False, + dropout, + temb_channels=512, + ): + super().__init__() + self.in_channels = in_channels + out_channels = in_channels if out_channels is None else out_channels + self.out_channels = out_channels + self.use_conv_shortcut = conv_shortcut + + self.norm1 = Normalize(in_channels) + self.conv1 = torch.nn.Conv2d( + in_channels, out_channels, kernel_size=3, stride=1, padding=1 + ) + if temb_channels > 0: + self.temb_proj = torch.nn.Linear(temb_channels, out_channels) + self.norm2 = Normalize(out_channels) + self.dropout = torch.nn.Dropout(dropout) + self.conv2 = torch.nn.Conv2d( + out_channels, out_channels, kernel_size=3, stride=1, padding=1 + ) + if self.in_channels != self.out_channels: + if self.use_conv_shortcut: + self.conv_shortcut = torch.nn.Conv2d( + in_channels, out_channels, kernel_size=3, stride=1, padding=1 + ) + else: + self.nin_shortcut = torch.nn.Conv2d( + in_channels, out_channels, kernel_size=1, stride=1, padding=0 + ) + + def forward(self, x, temb): + h = x + h = self.norm1(h) + h = nonlinearity(h) + h = self.conv1(h) + + if temb is not None: + h = h + self.temb_proj(nonlinearity(temb))[:, :, None, None] + + h = self.norm2(h) + h = nonlinearity(h) + h = self.dropout(h) + h = self.conv2(h) + + if self.in_channels != self.out_channels: + if self.use_conv_shortcut: + x = self.conv_shortcut(x) + else: + x = self.nin_shortcut(x) + + return x + h + + +class LinAttnBlock(LinearAttention): + """to match AttnBlock usage""" + + def __init__(self, in_channels): + super().__init__(dim=in_channels, heads=1, dim_head=in_channels) + + +class AttnBlock(nn.Module): + def __init__(self, in_channels): + super().__init__() + self.in_channels = in_channels + + self.norm = Normalize(in_channels) + self.q = torch.nn.Conv2d( + in_channels, in_channels, kernel_size=1, stride=1, padding=0 + ) + self.k = torch.nn.Conv2d( + in_channels, in_channels, kernel_size=1, stride=1, padding=0 + ) + self.v = torch.nn.Conv2d( + in_channels, in_channels, kernel_size=1, stride=1, padding=0 + ) + self.proj_out = torch.nn.Conv2d( + in_channels, in_channels, kernel_size=1, stride=1, padding=0 + ) + + def attention(self, h_: torch.Tensor) -> torch.Tensor: + h_ = self.norm(h_) + q = self.q(h_) + k = self.k(h_) + v = self.v(h_) + + b, c, h, w = q.shape + q, k, v = map( + lambda x: rearrange(x, "b c h w -> b 1 (h w) c").contiguous(), (q, k, v) + ) + h_ = torch.nn.functional.scaled_dot_product_attention( + q, k, v + ) # scale is dim ** -0.5 per default + # compute attention + + return rearrange(h_, "b 1 (h w) c -> b c h w", h=h, w=w, c=c, b=b) + + def forward(self, x, **kwargs): + h_ = x + h_ = self.attention(h_) + h_ = self.proj_out(h_) + return x + h_ + + +class MemoryEfficientAttnBlock(nn.Module): + """ + Uses xformers efficient implementation, + see https://github.com/MatthieuTPHR/diffusers/blob/d80b531ff8060ec1ea982b65a1b8df70f73aa67c/src/diffusers/models/attention.py#L223 + Note: this is a single-head self-attention operation + """ + + # + def __init__(self, in_channels): + super().__init__() + self.in_channels = in_channels + + self.norm = Normalize(in_channels) + self.q = torch.nn.Conv2d( + in_channels, in_channels, kernel_size=1, stride=1, padding=0 + ) + self.k = torch.nn.Conv2d( + in_channels, in_channels, kernel_size=1, stride=1, padding=0 + ) + self.v = torch.nn.Conv2d( + in_channels, in_channels, kernel_size=1, stride=1, padding=0 + ) + self.proj_out = torch.nn.Conv2d( + in_channels, in_channels, kernel_size=1, stride=1, padding=0 + ) + self.attention_op: Optional[Any] = None + + def attention(self, h_: torch.Tensor) -> torch.Tensor: + h_ = self.norm(h_) + q = self.q(h_) + k = self.k(h_) + v = self.v(h_) + + # compute attention + B, C, H, W = q.shape + q, k, v = map(lambda x: rearrange(x, "b c h w -> b (h w) c"), (q, k, v)) + + q, k, v = map( + lambda t: t.unsqueeze(3) + .reshape(B, t.shape[1], 1, C) + .permute(0, 2, 1, 3) + .reshape(B * 1, t.shape[1], C) + .contiguous(), + (q, k, v), + ) + out = xformers.ops.memory_efficient_attention( + q, k, v, attn_bias=None, op=self.attention_op + ) + + out = ( + out.unsqueeze(0) + .reshape(B, 1, out.shape[1], C) + .permute(0, 2, 1, 3) + .reshape(B, out.shape[1], C) + ) + return rearrange(out, "b (h w) c -> b c h w", b=B, h=H, w=W, c=C) + + def forward(self, x, **kwargs): + h_ = x + h_ = self.attention(h_) + h_ = self.proj_out(h_) + return x + h_ + + +class MemoryEfficientCrossAttentionWrapper(MemoryEfficientCrossAttention): + def forward(self, x, context=None, mask=None, **unused_kwargs): + b, c, h, w = x.shape + x = rearrange(x, "b c h w -> b (h w) c") + out = super().forward(x, context=context, mask=mask) + out = rearrange(out, "b (h w) c -> b c h w", h=h, w=w, c=c) + return x + out + + +def make_attn(in_channels, attn_type="vanilla", attn_kwargs=None): + assert attn_type in [ + "vanilla", + "vanilla-xformers", + "memory-efficient-cross-attn", + "linear", + "none", + ], f"attn_type {attn_type} unknown" + if ( + version.parse(torch.__version__) < version.parse("2.0.0") + and attn_type != "none" + ): + assert XFORMERS_IS_AVAILABLE, ( + f"We do not support vanilla attention in {torch.__version__} anymore, " + f"as it is too expensive. Please install xformers via e.g. 'pip install xformers==0.0.16'" + ) + attn_type = "vanilla-xformers" + logpy.info(f"making attention of type '{attn_type}' with {in_channels} in_channels") + if attn_type == "vanilla": + assert attn_kwargs is None + return AttnBlock(in_channels) + elif attn_type == "vanilla-xformers": + logpy.info( + f"building MemoryEfficientAttnBlock with {in_channels} in_channels..." + ) + return MemoryEfficientAttnBlock(in_channels) + elif type == "memory-efficient-cross-attn": + attn_kwargs["query_dim"] = in_channels + return MemoryEfficientCrossAttentionWrapper(**attn_kwargs) + elif attn_type == "none": + return nn.Identity(in_channels) + else: + return LinAttnBlock(in_channels) + + +class Model(nn.Module): + def __init__( + self, + *, + ch, + out_ch, + ch_mult=(1, 2, 4, 8), + num_res_blocks, + attn_resolutions, + dropout=0.0, + resamp_with_conv=True, + in_channels, + resolution, + use_timestep=True, + use_linear_attn=False, + attn_type="vanilla", + ): + super().__init__() + if use_linear_attn: + attn_type = "linear" + self.ch = ch + self.temb_ch = self.ch * 4 + self.num_resolutions = len(ch_mult) + self.num_res_blocks = num_res_blocks + self.resolution = resolution + self.in_channels = in_channels + + self.use_timestep = use_timestep + if self.use_timestep: + # timestep embedding + self.temb = nn.Module() + self.temb.dense = nn.ModuleList( + [ + torch.nn.Linear(self.ch, self.temb_ch), + torch.nn.Linear(self.temb_ch, self.temb_ch), + ] + ) + + # downsampling + self.conv_in = torch.nn.Conv2d( + in_channels, self.ch, kernel_size=3, stride=1, padding=1 + ) + + curr_res = resolution + in_ch_mult = (1,) + tuple(ch_mult) + self.down = nn.ModuleList() + for i_level in range(self.num_resolutions): + block = nn.ModuleList() + attn = nn.ModuleList() + block_in = ch * in_ch_mult[i_level] + block_out = ch * ch_mult[i_level] + for i_block in range(self.num_res_blocks): + block.append( + ResnetBlock( + in_channels=block_in, + out_channels=block_out, + temb_channels=self.temb_ch, + dropout=dropout, + ) + ) + block_in = block_out + if curr_res in attn_resolutions: + attn.append(make_attn(block_in, attn_type=attn_type)) + down = nn.Module() + down.block = block + down.attn = attn + if i_level != self.num_resolutions - 1: + down.downsample = Downsample(block_in, resamp_with_conv) + curr_res = curr_res // 2 + self.down.append(down) + + # middle + self.mid = nn.Module() + self.mid.block_1 = ResnetBlock( + in_channels=block_in, + out_channels=block_in, + temb_channels=self.temb_ch, + dropout=dropout, + ) + self.mid.attn_1 = make_attn(block_in, attn_type=attn_type) + self.mid.block_2 = ResnetBlock( + in_channels=block_in, + out_channels=block_in, + temb_channels=self.temb_ch, + dropout=dropout, + ) + + # upsampling + self.up = nn.ModuleList() + for i_level in reversed(range(self.num_resolutions)): + block = nn.ModuleList() + attn = nn.ModuleList() + block_out = ch * ch_mult[i_level] + skip_in = ch * ch_mult[i_level] + for i_block in range(self.num_res_blocks + 1): + if i_block == self.num_res_blocks: + skip_in = ch * in_ch_mult[i_level] + block.append( + ResnetBlock( + in_channels=block_in + skip_in, + out_channels=block_out, + temb_channels=self.temb_ch, + dropout=dropout, + ) + ) + block_in = block_out + if curr_res in attn_resolutions: + attn.append(make_attn(block_in, attn_type=attn_type)) + up = nn.Module() + up.block = block + up.attn = attn + if i_level != 0: + up.upsample = Upsample(block_in, resamp_with_conv) + curr_res = curr_res * 2 + self.up.insert(0, up) # prepend to get consistent order + + # end + self.norm_out = Normalize(block_in) + self.conv_out = torch.nn.Conv2d( + block_in, out_ch, kernel_size=3, stride=1, padding=1 + ) + + def forward(self, x, t=None, context=None): + # assert x.shape[2] == x.shape[3] == self.resolution + if context is not None: + # assume aligned context, cat along channel axis + x = torch.cat((x, context), dim=1) + if self.use_timestep: + # timestep embedding + assert t is not None + temb = get_timestep_embedding(t, self.ch) + temb = self.temb.dense[0](temb) + temb = nonlinearity(temb) + temb = self.temb.dense[1](temb) + else: + temb = None + + # downsampling + hs = [self.conv_in(x)] + for i_level in range(self.num_resolutions): + for i_block in range(self.num_res_blocks): + h = self.down[i_level].block[i_block](hs[-1], temb) + if len(self.down[i_level].attn) > 0: + h = self.down[i_level].attn[i_block](h) + hs.append(h) + if i_level != self.num_resolutions - 1: + hs.append(self.down[i_level].downsample(hs[-1])) + + # middle + h = hs[-1] + h = self.mid.block_1(h, temb) + h = self.mid.attn_1(h) + h = self.mid.block_2(h, temb) + + # upsampling + for i_level in reversed(range(self.num_resolutions)): + for i_block in range(self.num_res_blocks + 1): + h = self.up[i_level].block[i_block]( + torch.cat([h, hs.pop()], dim=1), temb + ) + if len(self.up[i_level].attn) > 0: + h = self.up[i_level].attn[i_block](h) + if i_level != 0: + h = self.up[i_level].upsample(h) + + # end + h = self.norm_out(h) + h = nonlinearity(h) + h = self.conv_out(h) + return h + + def get_last_layer(self): + return self.conv_out.weight + + +class Encoder(nn.Module): + def __init__( + self, + *, + ch, + out_ch, + ch_mult=(1, 2, 4, 8), + num_res_blocks, + attn_resolutions, + dropout=0.0, + resamp_with_conv=True, + in_channels, + resolution, + z_channels, + double_z=True, + use_linear_attn=False, + attn_type="vanilla", + **ignore_kwargs, + ): + super().__init__() + if use_linear_attn: + attn_type = "linear" + self.ch = ch + self.temb_ch = 0 + self.num_resolutions = len(ch_mult) + self.num_res_blocks = num_res_blocks + self.resolution = resolution + self.in_channels = in_channels + + # downsampling + self.conv_in = torch.nn.Conv2d( + in_channels, self.ch, kernel_size=3, stride=1, padding=1 + ) + + curr_res = resolution + in_ch_mult = (1,) + tuple(ch_mult) + self.in_ch_mult = in_ch_mult + self.down = nn.ModuleList() + for i_level in range(self.num_resolutions): + block = nn.ModuleList() + attn = nn.ModuleList() + block_in = ch * in_ch_mult[i_level] + block_out = ch * ch_mult[i_level] + for i_block in range(self.num_res_blocks): + block.append( + ResnetBlock( + in_channels=block_in, + out_channels=block_out, + temb_channels=self.temb_ch, + dropout=dropout, + ) + ) + block_in = block_out + if curr_res in attn_resolutions: + attn.append(make_attn(block_in, attn_type=attn_type)) + down = nn.Module() + down.block = block + down.attn = attn + if i_level != self.num_resolutions - 1: + down.downsample = Downsample(block_in, resamp_with_conv) + curr_res = curr_res // 2 + self.down.append(down) + + # middle + self.mid = nn.Module() + self.mid.block_1 = ResnetBlock( + in_channels=block_in, + out_channels=block_in, + temb_channels=self.temb_ch, + dropout=dropout, + ) + self.mid.attn_1 = make_attn(block_in, attn_type=attn_type) + self.mid.block_2 = ResnetBlock( + in_channels=block_in, + out_channels=block_in, + temb_channels=self.temb_ch, + dropout=dropout, + ) + + # end + self.norm_out = Normalize(block_in) + self.conv_out = torch.nn.Conv2d( + block_in, + 2 * z_channels if double_z else z_channels, + kernel_size=3, + stride=1, + padding=1, + ) + + def forward(self, x): + # timestep embedding + temb = None + + # downsampling + hs = [self.conv_in(x)] + for i_level in range(self.num_resolutions): + for i_block in range(self.num_res_blocks): + h = self.down[i_level].block[i_block](hs[-1], temb) + if len(self.down[i_level].attn) > 0: + h = self.down[i_level].attn[i_block](h) + hs.append(h) + if i_level != self.num_resolutions - 1: + hs.append(self.down[i_level].downsample(hs[-1])) + + # middle + h = hs[-1] + h = self.mid.block_1(h, temb) + h = self.mid.attn_1(h) + h = self.mid.block_2(h, temb) + + # end + h = self.norm_out(h) + h = nonlinearity(h) + h = self.conv_out(h) + return h + + +class Decoder(nn.Module): + def __init__( + self, + *, + ch, + out_ch, + ch_mult=(1, 2, 4, 8), + num_res_blocks, + attn_resolutions, + dropout=0.0, + resamp_with_conv=True, + in_channels, + resolution, + z_channels, + give_pre_end=False, + tanh_out=False, + use_linear_attn=False, + attn_type="vanilla", + **ignorekwargs, + ): + super().__init__() + if use_linear_attn: + attn_type = "linear" + self.ch = ch + self.temb_ch = 0 + self.num_resolutions = len(ch_mult) + self.num_res_blocks = num_res_blocks + self.resolution = resolution + self.in_channels = in_channels + self.give_pre_end = give_pre_end + self.tanh_out = tanh_out + + # compute in_ch_mult, block_in and curr_res at lowest res + in_ch_mult = (1,) + tuple(ch_mult) + block_in = ch * ch_mult[self.num_resolutions - 1] + curr_res = resolution // 2 ** (self.num_resolutions - 1) + self.z_shape = (1, z_channels, curr_res, curr_res) + logpy.info( + "Working with z of shape {} = {} dimensions.".format( + self.z_shape, np.prod(self.z_shape) + ) + ) + + make_attn_cls = self._make_attn() + make_resblock_cls = self._make_resblock() + make_conv_cls = self._make_conv() + # z to block_in + self.conv_in = torch.nn.Conv2d( + z_channels, block_in, kernel_size=3, stride=1, padding=1 + ) + + # middle + self.mid = nn.Module() + self.mid.block_1 = make_resblock_cls( + in_channels=block_in, + out_channels=block_in, + temb_channels=self.temb_ch, + dropout=dropout, + ) + self.mid.attn_1 = make_attn_cls(block_in, attn_type=attn_type) + self.mid.block_2 = make_resblock_cls( + in_channels=block_in, + out_channels=block_in, + temb_channels=self.temb_ch, + dropout=dropout, + ) + + # upsampling + self.up = nn.ModuleList() + for i_level in reversed(range(self.num_resolutions)): + block = nn.ModuleList() + attn = nn.ModuleList() + block_out = ch * ch_mult[i_level] + for i_block in range(self.num_res_blocks + 1): + block.append( + make_resblock_cls( + in_channels=block_in, + out_channels=block_out, + temb_channels=self.temb_ch, + dropout=dropout, + ) + ) + block_in = block_out + if curr_res in attn_resolutions: + attn.append(make_attn_cls(block_in, attn_type=attn_type)) + up = nn.Module() + up.block = block + up.attn = attn + if i_level != 0: + up.upsample = Upsample(block_in, resamp_with_conv) + curr_res = curr_res * 2 + self.up.insert(0, up) # prepend to get consistent order + + # end + self.norm_out = Normalize(block_in) + self.conv_out = make_conv_cls( + block_in, out_ch, kernel_size=3, stride=1, padding=1 + ) + + def _make_attn(self) -> Callable: + return make_attn + + def _make_resblock(self) -> Callable: + return ResnetBlock + + def _make_conv(self) -> Callable: + return torch.nn.Conv2d + + def get_last_layer(self, **kwargs): + return self.conv_out.weight + + def forward(self, z, **kwargs): + # assert z.shape[1:] == self.z_shape[1:] + self.last_z_shape = z.shape + + # timestep embedding + temb = None + + # z to block_in + h = self.conv_in(z) + + # middle + h = self.mid.block_1(h, temb, **kwargs) + h = self.mid.attn_1(h, **kwargs) + h = self.mid.block_2(h, temb, **kwargs) + + # upsampling + for i_level in reversed(range(self.num_resolutions)): + for i_block in range(self.num_res_blocks + 1): + h = self.up[i_level].block[i_block](h, temb, **kwargs) + if len(self.up[i_level].attn) > 0: + h = self.up[i_level].attn[i_block](h, **kwargs) + if i_level != 0: + h = self.up[i_level].upsample(h) + + # end + if self.give_pre_end: + return h + + h = self.norm_out(h) + h = nonlinearity(h) + h = self.conv_out(h, **kwargs) + if self.tanh_out: + h = torch.tanh(h) + return h diff --git a/sgm/modules/diffusionmodules/openaimodel.py b/sgm/modules/diffusionmodules/openaimodel.py new file mode 100644 index 0000000..275f4a0 --- /dev/null +++ b/sgm/modules/diffusionmodules/openaimodel.py @@ -0,0 +1,885 @@ +import logging +import math +from abc import abstractmethod +from typing import Iterable, List, Optional, Tuple, Union + +import torch as th +import torch.nn as nn +import torch.nn.functional as F +from einops import rearrange +from torch.utils.checkpoint import checkpoint + +from ...modules.attention import SpatialTransformer +from ...modules.diffusionmodules.util import (avg_pool_nd, conv_nd, linear, + normalization, + timestep_embedding, zero_module) +from ...modules.video_attention import SpatialVideoTransformer +from ...util import exists + +logpy = logging.getLogger(__name__) + + +class AttentionPool2d(nn.Module): + """ + Adapted from CLIP: https://github.com/openai/CLIP/blob/main/clip/model.py + """ + + def __init__( + self, + spacial_dim: int, + embed_dim: int, + num_heads_channels: int, + output_dim: Optional[int] = None, + ): + super().__init__() + self.positional_embedding = nn.Parameter( + th.randn(embed_dim, spacial_dim**2 + 1) / embed_dim**0.5 + ) + self.qkv_proj = conv_nd(1, embed_dim, 3 * embed_dim, 1) + self.c_proj = conv_nd(1, embed_dim, output_dim or embed_dim, 1) + self.num_heads = embed_dim // num_heads_channels + self.attention = QKVAttention(self.num_heads) + + def forward(self, x: th.Tensor) -> th.Tensor: + b, c, _ = x.shape + x = x.reshape(b, c, -1) + x = th.cat([x.mean(dim=-1, keepdim=True), x], dim=-1) + x = x + self.positional_embedding[None, :, :].to(x.dtype) + x = self.qkv_proj(x) + x = self.attention(x) + x = self.c_proj(x) + return x[:, :, 0] + + +class TimestepBlock(nn.Module): + """ + Any module where forward() takes timestep embeddings as a second argument. + """ + + @abstractmethod + def forward(self, x: th.Tensor, emb: th.Tensor): + """ + Apply the module to `x` given `emb` timestep embeddings. + """ + + +class TimestepEmbedSequential(nn.Sequential, TimestepBlock): + """ + A sequential module that passes timestep embeddings to the children that + support it as an extra input. + """ + + def forward( + self, + x: th.Tensor, + emb: th.Tensor, + context: Optional[th.Tensor] = None, + image_only_indicator: Optional[th.Tensor] = None, + cond_view: Optional[th.Tensor] = None, + cond_motion: Optional[th.Tensor] = None, + time_context: Optional[int] = None, + num_video_frames: Optional[int] = None, + time_step: Optional[int] = None, + name: Optional[str] = None, + ): + from ...modules.diffusionmodules.video_model import VideoResBlock, PostHocResBlockWithTime + from ...modules.spacetime_attention import ( + BasicTransformerTimeMixBlock, + PostHocSpatialTransformerWithTimeMixing, + PostHocSpatialTransformerWithTimeMixingAndMotion + ) + + for layer in self: + module = layer + + if isinstance( + module, + ( + BasicTransformerTimeMixBlock, + PostHocSpatialTransformerWithTimeMixing, + PostHocSpatialTransformerWithTimeMixingAndMotion + ), + ): + x = layer( + x, + context, + # cam, + time_context, + num_video_frames, + image_only_indicator, + cond_view, + cond_motion, + time_step, + name, + ) + elif isinstance(module, SpatialVideoTransformer): + x = layer( + x, + context, + time_context, + num_video_frames, + image_only_indicator, + # time_step, + ) + elif isinstance(module, PostHocResBlockWithTime): + x = layer(x, emb, num_video_frames, image_only_indicator) + elif isinstance(module, VideoResBlock): + x = layer(x, emb, num_video_frames, image_only_indicator) + elif isinstance(module, TimestepBlock) and not isinstance( + module, VideoResBlock + ): + x = layer(x, emb) + elif isinstance(module, SpatialTransformer): + x = layer(x, context) + else: + x = layer(x) + return x + + +class Upsample(nn.Module): + """ + An upsampling layer with an optional convolution. + :param channels: channels in the inputs and outputs. + :param use_conv: a bool determining if a convolution is applied. + :param dims: determines if the signal is 1D, 2D, or 3D. If 3D, then + upsampling occurs in the inner-two dimensions. + """ + + def __init__( + self, + channels: int, + use_conv: bool, + dims: int = 2, + out_channels: Optional[int] = None, + padding: int = 1, + third_up: bool = False, + kernel_size: int = 3, + scale_factor: int = 2, + ): + super().__init__() + self.channels = channels + self.out_channels = out_channels or channels + self.use_conv = use_conv + self.dims = dims + self.third_up = third_up + self.scale_factor = scale_factor + if use_conv: + self.conv = conv_nd( + dims, self.channels, self.out_channels, kernel_size, padding=padding + ) + + def forward(self, x: th.Tensor) -> th.Tensor: + assert x.shape[1] == self.channels + + if self.dims == 3: + t_factor = 1 if not self.third_up else self.scale_factor + x = F.interpolate( + x, + ( + t_factor * x.shape[2], + x.shape[3] * self.scale_factor, + x.shape[4] * self.scale_factor, + ), + mode="nearest", + ) + else: + x = F.interpolate(x, scale_factor=self.scale_factor, mode="nearest") + if self.use_conv: + x = self.conv(x) + return x + + +class Downsample(nn.Module): + """ + A downsampling layer with an optional convolution. + :param channels: channels in the inputs and outputs. + :param use_conv: a bool determining if a convolution is applied. + :param dims: determines if the signal is 1D, 2D, or 3D. If 3D, then + downsampling occurs in the inner-two dimensions. + """ + + def __init__( + self, + channels: int, + use_conv: bool, + dims: int = 2, + out_channels: Optional[int] = None, + padding: int = 1, + third_down: bool = False, + ): + super().__init__() + self.channels = channels + self.out_channels = out_channels or channels + self.use_conv = use_conv + self.dims = dims + stride = 2 if dims != 3 else ((1, 2, 2) if not third_down else (2, 2, 2)) + if use_conv: + logpy.info(f"Building a Downsample layer with {dims} dims.") + logpy.info( + f" --> settings are: \n in-chn: {self.channels}, out-chn: {self.out_channels}, " + f"kernel-size: 3, stride: {stride}, padding: {padding}" + ) + if dims == 3: + logpy.info(f" --> Downsampling third axis (time): {third_down}") + self.op = conv_nd( + dims, + self.channels, + self.out_channels, + 3, + stride=stride, + padding=padding, + ) + else: + assert self.channels == self.out_channels + self.op = avg_pool_nd(dims, kernel_size=stride, stride=stride) + + def forward(self, x: th.Tensor) -> th.Tensor: + assert x.shape[1] == self.channels + + return self.op(x) + + +class ResBlock(TimestepBlock): + """ + A residual block that can optionally change the number of channels. + :param channels: the number of input channels. + :param emb_channels: the number of timestep embedding channels. + :param dropout: the rate of dropout. + :param out_channels: if specified, the number of out channels. + :param use_conv: if True and out_channels is specified, use a spatial + convolution instead of a smaller 1x1 convolution to change the + channels in the skip connection. + :param dims: determines if the signal is 1D, 2D, or 3D. + :param use_checkpoint: if True, use gradient checkpointing on this module. + :param up: if True, use this block for upsampling. + :param down: if True, use this block for downsampling. + """ + + def __init__( + self, + channels: int, + emb_channels: int, + dropout: float, + out_channels: Optional[int] = None, + use_conv: bool = False, + use_scale_shift_norm: bool = False, + dims: int = 2, + use_checkpoint: bool = False, + up: bool = False, + down: bool = False, + kernel_size: int = 3, + exchange_temb_dims: bool = False, + skip_t_emb: bool = False, + ): + super().__init__() + self.channels = channels + self.emb_channels = emb_channels + self.dropout = dropout + self.out_channels = out_channels or channels + self.use_conv = use_conv + self.use_checkpoint = use_checkpoint + self.use_scale_shift_norm = use_scale_shift_norm + self.exchange_temb_dims = exchange_temb_dims + + if isinstance(kernel_size, Iterable): + padding = [k // 2 for k in kernel_size] + else: + padding = kernel_size // 2 + + self.in_layers = nn.Sequential( + normalization(channels), + nn.SiLU(), + conv_nd(dims, channels, self.out_channels, kernel_size, padding=padding), + ) + + self.updown = up or down + + if up: + self.h_upd = Upsample(channels, False, dims) + self.x_upd = Upsample(channels, False, dims) + elif down: + self.h_upd = Downsample(channels, False, dims) + self.x_upd = Downsample(channels, False, dims) + else: + self.h_upd = self.x_upd = nn.Identity() + + self.skip_t_emb = skip_t_emb + self.emb_out_channels = ( + 2 * self.out_channels if use_scale_shift_norm else self.out_channels + ) + if self.skip_t_emb: + logpy.info(f"Skipping timestep embedding in {self.__class__.__name__}") + assert not self.use_scale_shift_norm + self.emb_layers = None + self.exchange_temb_dims = False + else: + self.emb_layers = nn.Sequential( + nn.SiLU(), + linear( + emb_channels, + self.emb_out_channels, + ), + ) + + self.out_layers = nn.Sequential( + normalization(self.out_channels), + nn.SiLU(), + nn.Dropout(p=dropout), + zero_module( + conv_nd( + dims, + self.out_channels, + self.out_channels, + kernel_size, + padding=padding, + ) + ), + ) + + if self.out_channels == channels: + self.skip_connection = nn.Identity() + elif use_conv: + self.skip_connection = conv_nd( + dims, channels, self.out_channels, kernel_size, padding=padding + ) + else: + self.skip_connection = conv_nd(dims, channels, self.out_channels, 1) + + def forward(self, x: th.Tensor, emb: th.Tensor) -> th.Tensor: + """ + Apply the block to a Tensor, conditioned on a timestep embedding. + :param x: an [N x C x ...] Tensor of features. + :param emb: an [N x emb_channels] Tensor of timestep embeddings. + :return: an [N x C x ...] Tensor of outputs. + """ + if self.use_checkpoint: + return checkpoint(self._forward, x, emb) + else: + return self._forward(x, emb) + + def _forward(self, x: th.Tensor, emb: th.Tensor) -> th.Tensor: + if self.updown: + in_rest, in_conv = self.in_layers[:-1], self.in_layers[-1] + h = in_rest(x) + h = self.h_upd(h) + x = self.x_upd(x) + h = in_conv(h) + else: + h = self.in_layers(x) + + if self.skip_t_emb: + emb_out = th.zeros_like(h) + else: + emb_out = self.emb_layers(emb).type(h.dtype) + while len(emb_out.shape) < len(h.shape): + emb_out = emb_out[..., None] + if self.use_scale_shift_norm: + out_norm, out_rest = self.out_layers[0], self.out_layers[1:] + scale, shift = th.chunk(emb_out, 2, dim=1) + h = out_norm(h) * (1 + scale) + shift + h = out_rest(h) + else: + if self.exchange_temb_dims: + emb_out = rearrange(emb_out, "b t c ... -> b c t ...") + h = h + emb_out + h = self.out_layers(h) + return self.skip_connection(x) + h + + +class AttentionBlock(nn.Module): + """ + An attention block that allows spatial positions to attend to each other. + Originally ported from here, but adapted to the N-d case. + https://github.com/hojonathanho/diffusion/blob/1e0dceb3b3495bbe19116a5e1b3596cd0706c543/diffusion_tf/models/unet.py#L66. + """ + + def __init__( + self, + channels: int, + num_heads: int = 1, + num_head_channels: int = -1, + use_checkpoint: bool = False, + use_new_attention_order: bool = False, + ): + super().__init__() + self.channels = channels + if num_head_channels == -1: + self.num_heads = num_heads + else: + assert ( + channels % num_head_channels == 0 + ), f"q,k,v channels {channels} is not divisible by num_head_channels {num_head_channels}" + self.num_heads = channels // num_head_channels + self.use_checkpoint = use_checkpoint + self.norm = normalization(channels) + self.qkv = conv_nd(1, channels, channels * 3, 1) + if use_new_attention_order: + # split qkv before split heads + self.attention = QKVAttention(self.num_heads) + else: + # split heads before split qkv + self.attention = QKVAttentionLegacy(self.num_heads) + + self.proj_out = zero_module(conv_nd(1, channels, channels, 1)) + + def forward(self, x: th.Tensor, **kwargs) -> th.Tensor: + return checkpoint(self._forward, x) + + def _forward(self, x: th.Tensor) -> th.Tensor: + b, c, *spatial = x.shape + x = x.reshape(b, c, -1) + qkv = self.qkv(self.norm(x)) + h = self.attention(qkv) + h = self.proj_out(h) + return (x + h).reshape(b, c, *spatial) + + +class QKVAttentionLegacy(nn.Module): + """ + A module which performs QKV attention. Matches legacy QKVAttention + input/ouput heads shaping + """ + + def __init__(self, n_heads: int): + super().__init__() + self.n_heads = n_heads + + def forward(self, qkv: th.Tensor) -> th.Tensor: + """ + Apply QKV attention. + :param qkv: an [N x (H * 3 * C) x T] tensor of Qs, Ks, and Vs. + :return: an [N x (H * C) x T] tensor after attention. + """ + bs, width, length = qkv.shape + assert width % (3 * self.n_heads) == 0 + ch = width // (3 * self.n_heads) + q, k, v = qkv.reshape(bs * self.n_heads, ch * 3, length).split(ch, dim=1) + scale = 1 / math.sqrt(math.sqrt(ch)) + weight = th.einsum( + "bct,bcs->bts", q * scale, k * scale + ) # More stable with f16 than dividing afterwards + weight = th.softmax(weight.float(), dim=-1).type(weight.dtype) + a = th.einsum("bts,bcs->bct", weight, v) + return a.reshape(bs, -1, length) + + +class QKVAttention(nn.Module): + """ + A module which performs QKV attention and splits in a different order. + """ + + def __init__(self, n_heads: int): + super().__init__() + self.n_heads = n_heads + + def forward(self, qkv: th.Tensor) -> th.Tensor: + """ + Apply QKV attention. + :param qkv: an [N x (3 * H * C) x T] tensor of Qs, Ks, and Vs. + :return: an [N x (H * C) x T] tensor after attention. + """ + bs, width, length = qkv.shape + assert width % (3 * self.n_heads) == 0 + ch = width // (3 * self.n_heads) + q, k, v = qkv.chunk(3, dim=1) + scale = 1 / math.sqrt(math.sqrt(ch)) + weight = th.einsum( + "bct,bcs->bts", + (q * scale).view(bs * self.n_heads, ch, length), + (k * scale).view(bs * self.n_heads, ch, length), + ) # More stable with f16 than dividing afterwards + weight = th.softmax(weight.float(), dim=-1).type(weight.dtype) + a = th.einsum("bts,bcs->bct", weight, v.reshape(bs * self.n_heads, ch, length)) + return a.reshape(bs, -1, length) + + +class Timestep(nn.Module): + def __init__(self, dim: int): + super().__init__() + self.dim = dim + + def forward(self, t: th.Tensor) -> th.Tensor: + return timestep_embedding(t, self.dim) + + +class UNetModel(nn.Module): + """ + The full UNet model with attention and timestep embedding. + :param in_channels: channels in the input Tensor. + :param model_channels: base channel count for the model. + :param out_channels: channels in the output Tensor. + :param num_res_blocks: number of residual blocks per downsample. + :param attention_resolutions: a collection of downsample rates at which + attention will take place. May be a set, list, or tuple. + For example, if this contains 4, then at 4x downsampling, attention + will be used. + :param dropout: the dropout probability. + :param channel_mult: channel multiplier for each level of the UNet. + :param conv_resample: if True, use learned convolutions for upsampling and + downsampling. + :param dims: determines if the signal is 1D, 2D, or 3D. + :param num_classes: if specified (as an int), then this model will be + class-conditional with `num_classes` classes. + :param use_checkpoint: use gradient checkpointing to reduce memory usage. + :param num_heads: the number of attention heads in each attention layer. + :param num_heads_channels: if specified, ignore num_heads and instead use + a fixed channel width per attention head. + :param num_heads_upsample: works with num_heads to set a different number + of heads for upsampling. Deprecated. + :param use_scale_shift_norm: use a FiLM-like conditioning mechanism. + :param resblock_updown: use residual blocks for up/downsampling. + :param use_new_attention_order: use a different attention pattern for potentially + increased efficiency. + """ + + def __init__( + self, + in_channels: int, + model_channels: int, + out_channels: int, + num_res_blocks: int, + attention_resolutions: int, + dropout: float = 0.0, + channel_mult: Union[List, Tuple] = (1, 2, 4, 8), + conv_resample: bool = True, + dims: int = 2, + num_classes: Optional[Union[int, str]] = None, + use_checkpoint: bool = False, + num_heads: int = -1, + num_head_channels: int = -1, + num_heads_upsample: int = -1, + use_scale_shift_norm: bool = False, + resblock_updown: bool = False, + transformer_depth: int = 1, + context_dim: Optional[int] = None, + disable_self_attentions: Optional[List[bool]] = None, + num_attention_blocks: Optional[List[int]] = None, + disable_middle_self_attn: bool = False, + disable_middle_transformer: bool = False, + use_linear_in_transformer: bool = False, + spatial_transformer_attn_type: str = "softmax", + adm_in_channels: Optional[int] = None, + ): + super().__init__() + + if num_heads_upsample == -1: + num_heads_upsample = num_heads + + if num_heads == -1: + assert ( + num_head_channels != -1 + ), "Either num_heads or num_head_channels has to be set" + + if num_head_channels == -1: + assert ( + num_heads != -1 + ), "Either num_heads or num_head_channels has to be set" + + self.in_channels = in_channels + self.model_channels = model_channels + self.out_channels = out_channels + if isinstance(transformer_depth, int): + transformer_depth = len(channel_mult) * [transformer_depth] + transformer_depth_middle = transformer_depth[-1] + + if isinstance(num_res_blocks, int): + self.num_res_blocks = len(channel_mult) * [num_res_blocks] + else: + if len(num_res_blocks) != len(channel_mult): + raise ValueError( + "provide num_res_blocks either as an int (globally constant) or " + "as a list/tuple (per-level) with the same length as channel_mult" + ) + self.num_res_blocks = num_res_blocks + + if disable_self_attentions is not None: + assert len(disable_self_attentions) == len(channel_mult) + if num_attention_blocks is not None: + assert len(num_attention_blocks) == len(self.num_res_blocks) + assert all( + map( + lambda i: self.num_res_blocks[i] >= num_attention_blocks[i], + range(len(num_attention_blocks)), + ) + ) + logpy.info( + f"Constructor of UNetModel received num_attention_blocks={num_attention_blocks}. " + f"This option has LESS priority than attention_resolutions {attention_resolutions}, " + f"i.e., in cases where num_attention_blocks[i] > 0 but 2**i not in attention_resolutions, " + f"attention will still not be set." + ) + + self.attention_resolutions = attention_resolutions + self.dropout = dropout + self.channel_mult = channel_mult + self.conv_resample = conv_resample + self.num_classes = num_classes + self.use_checkpoint = use_checkpoint + self.num_heads = num_heads + self.num_head_channels = num_head_channels + self.num_heads_upsample = num_heads_upsample + + time_embed_dim = model_channels * 4 + self.time_embed = nn.Sequential( + linear(model_channels, time_embed_dim), + nn.SiLU(), + linear(time_embed_dim, time_embed_dim), + ) + + if self.num_classes is not None: + if isinstance(self.num_classes, int): + self.label_emb = nn.Embedding(num_classes, time_embed_dim) + elif self.num_classes == "continuous": + logpy.info("setting up linear c_adm embedding layer") + self.label_emb = nn.Linear(1, time_embed_dim) + elif self.num_classes == "timestep": + self.label_emb = nn.Sequential( + Timestep(model_channels), + nn.Sequential( + linear(model_channels, time_embed_dim), + nn.SiLU(), + linear(time_embed_dim, time_embed_dim), + ), + ) + elif self.num_classes == "sequential": + assert adm_in_channels is not None + self.label_emb = nn.Sequential( + nn.Sequential( + linear(adm_in_channels, time_embed_dim), + nn.SiLU(), + linear(time_embed_dim, time_embed_dim), + ) + ) + else: + raise ValueError + + self.input_blocks = nn.ModuleList( + [ + TimestepEmbedSequential( + conv_nd(dims, in_channels, model_channels, 3, padding=1) + ) + ] + ) + self._feature_size = model_channels + input_block_chans = [model_channels] + ch = model_channels + ds = 1 + for level, mult in enumerate(channel_mult): + for nr in range(self.num_res_blocks[level]): + layers = [ + ResBlock( + ch, + time_embed_dim, + dropout, + out_channels=mult * model_channels, + dims=dims, + use_checkpoint=use_checkpoint, + use_scale_shift_norm=use_scale_shift_norm, + ) + ] + ch = mult * model_channels + if ds in attention_resolutions: + if num_head_channels == -1: + dim_head = ch // num_heads + else: + num_heads = ch // num_head_channels + dim_head = num_head_channels + + if context_dim is not None and exists(disable_self_attentions): + disabled_sa = disable_self_attentions[level] + else: + disabled_sa = False + + if ( + not exists(num_attention_blocks) + or nr < num_attention_blocks[level] + ): + layers.append( + SpatialTransformer( + ch, + num_heads, + dim_head, + depth=transformer_depth[level], + context_dim=context_dim, + disable_self_attn=disabled_sa, + use_linear=use_linear_in_transformer, + attn_type=spatial_transformer_attn_type, + use_checkpoint=use_checkpoint, + ) + ) + self.input_blocks.append(TimestepEmbedSequential(*layers)) + self._feature_size += ch + input_block_chans.append(ch) + if level != len(channel_mult) - 1: + out_ch = ch + self.input_blocks.append( + TimestepEmbedSequential( + ResBlock( + ch, + time_embed_dim, + dropout, + out_channels=out_ch, + dims=dims, + use_checkpoint=use_checkpoint, + use_scale_shift_norm=use_scale_shift_norm, + down=True, + ) + if resblock_updown + else Downsample( + ch, conv_resample, dims=dims, out_channels=out_ch + ) + ) + ) + ch = out_ch + input_block_chans.append(ch) + ds *= 2 + self._feature_size += ch + + if num_head_channels == -1: + dim_head = ch // num_heads + else: + num_heads = ch // num_head_channels + dim_head = num_head_channels + + self.middle_block = TimestepEmbedSequential( + ResBlock( + ch, + time_embed_dim, + dropout, + out_channels=ch, + dims=dims, + use_checkpoint=use_checkpoint, + use_scale_shift_norm=use_scale_shift_norm, + ), + SpatialTransformer( + ch, + num_heads, + dim_head, + depth=transformer_depth_middle, + context_dim=context_dim, + disable_self_attn=disable_middle_self_attn, + use_linear=use_linear_in_transformer, + attn_type=spatial_transformer_attn_type, + use_checkpoint=use_checkpoint, + ) + if not disable_middle_transformer + else th.nn.Identity(), + ResBlock( + ch, + time_embed_dim, + dropout, + dims=dims, + use_checkpoint=use_checkpoint, + use_scale_shift_norm=use_scale_shift_norm, + ), + ) + self._feature_size += ch + + self.output_blocks = nn.ModuleList([]) + for level, mult in list(enumerate(channel_mult))[::-1]: + for i in range(self.num_res_blocks[level] + 1): + ich = input_block_chans.pop() + layers = [ + ResBlock( + ch + ich, + time_embed_dim, + dropout, + out_channels=model_channels * mult, + dims=dims, + use_checkpoint=use_checkpoint, + use_scale_shift_norm=use_scale_shift_norm, + ) + ] + ch = model_channels * mult + if ds in attention_resolutions: + if num_head_channels == -1: + dim_head = ch // num_heads + else: + num_heads = ch // num_head_channels + dim_head = num_head_channels + + if exists(disable_self_attentions): + disabled_sa = disable_self_attentions[level] + else: + disabled_sa = False + + if ( + not exists(num_attention_blocks) + or i < num_attention_blocks[level] + ): + layers.append( + SpatialTransformer( + ch, + num_heads, + dim_head, + depth=transformer_depth[level], + context_dim=context_dim, + disable_self_attn=disabled_sa, + use_linear=use_linear_in_transformer, + attn_type=spatial_transformer_attn_type, + use_checkpoint=use_checkpoint, + ) + ) + if level and i == self.num_res_blocks[level]: + out_ch = ch + layers.append( + ResBlock( + ch, + time_embed_dim, + dropout, + out_channels=out_ch, + dims=dims, + use_checkpoint=use_checkpoint, + use_scale_shift_norm=use_scale_shift_norm, + up=True, + ) + if resblock_updown + else Upsample(ch, conv_resample, dims=dims, out_channels=out_ch) + ) + ds //= 2 + self.output_blocks.append(TimestepEmbedSequential(*layers)) + self._feature_size += ch + + self.out = nn.Sequential( + normalization(ch), + nn.SiLU(), + zero_module(conv_nd(dims, model_channels, out_channels, 3, padding=1)), + ) + + def forward( + self, + x: th.Tensor, + timesteps: Optional[th.Tensor] = None, + context: Optional[th.Tensor] = None, + y: Optional[th.Tensor] = None, + **kwargs, + ) -> th.Tensor: + """ + Apply the model to an input batch. + :param x: an [N x C x ...] Tensor of inputs. + :param timesteps: a 1-D batch of timesteps. + :param context: conditioning plugged in via crossattn + :param y: an [N] Tensor of labels, if class-conditional. + :return: an [N x C x ...] Tensor of outputs. + """ + assert (y is not None) == ( + self.num_classes is not None + ), "must specify y if and only if the model is class-conditional" + hs = [] + t_emb = timestep_embedding(timesteps, self.model_channels, repeat_only=False) + emb = self.time_embed(t_emb) + + if self.num_classes is not None: + assert y.shape[0] == x.shape[0] + emb = emb + self.label_emb(y) + + h = x + for module in self.input_blocks: + h = module(h, emb, context) + hs.append(h) + h = self.middle_block(h, emb, context) + for module in self.output_blocks: + h = th.cat([h, hs.pop()], dim=1) + h = module(h, emb, context) + h = h.type(x.dtype) + + return self.out(h) diff --git a/sgm/modules/diffusionmodules/sampling.py b/sgm/modules/diffusionmodules/sampling.py new file mode 100644 index 0000000..af07566 --- /dev/null +++ b/sgm/modules/diffusionmodules/sampling.py @@ -0,0 +1,362 @@ +""" + Partially ported from https://github.com/crowsonkb/k-diffusion/blob/master/k_diffusion/sampling.py +""" + + +from typing import Dict, Union + +import torch +from omegaconf import ListConfig, OmegaConf +from tqdm import tqdm + +from ...modules.diffusionmodules.sampling_utils import (get_ancestral_step, + linear_multistep_coeff, + to_d, to_neg_log_sigma, + to_sigma) +from ...util import append_dims, default, instantiate_from_config + +DEFAULT_GUIDER = {"target": "sgm.modules.diffusionmodules.guiders.IdentityGuider"} + + +class BaseDiffusionSampler: + def __init__( + self, + discretization_config: Union[Dict, ListConfig, OmegaConf], + num_steps: Union[int, None] = None, + guider_config: Union[Dict, ListConfig, OmegaConf, None] = None, + verbose: bool = False, + device: str = "cuda", + ): + self.num_steps = num_steps + self.discretization = instantiate_from_config(discretization_config) + self.guider = instantiate_from_config( + default( + guider_config, + DEFAULT_GUIDER, + ) + ) + self.verbose = verbose + self.device = device + + def prepare_sampling_loop(self, x, cond, uc=None, num_steps=None): + sigmas = self.discretization( + self.num_steps if num_steps is None else num_steps, device=self.device + ) + uc = default(uc, cond) + + x *= torch.sqrt(1.0 + sigmas[0] ** 2.0) + num_sigmas = len(sigmas) + + s_in = x.new_ones([x.shape[0]]) + + return x, s_in, sigmas, num_sigmas, cond, uc + + def denoise(self, x, denoiser, sigma, cond, uc): + denoised = denoiser(*self.guider.prepare_inputs(x, sigma, cond, uc)) + denoised = self.guider(denoised, sigma) + return denoised + + def get_sigma_gen(self, num_sigmas): + sigma_generator = range(num_sigmas - 1) + if self.verbose: + print("#" * 30, " Sampling setting ", "#" * 30) + print(f"Sampler: {self.__class__.__name__}") + print(f"Discretization: {self.discretization.__class__.__name__}") + print(f"Guider: {self.guider.__class__.__name__}") + sigma_generator = tqdm( + sigma_generator, + total=num_sigmas, + desc=f"Sampling with {self.__class__.__name__} for {num_sigmas} steps", + ) + return sigma_generator + + +class SingleStepDiffusionSampler(BaseDiffusionSampler): + def sampler_step(self, sigma, next_sigma, denoiser, x, cond, uc, *args, **kwargs): + raise NotImplementedError + + def euler_step(self, x, d, dt): + return x + dt * d + + +class EDMSampler(SingleStepDiffusionSampler): + def __init__( + self, s_churn=0.0, s_tmin=0.0, s_tmax=float("inf"), s_noise=1.0, *args, **kwargs + ): + super().__init__(*args, **kwargs) + + self.s_churn = s_churn + self.s_tmin = s_tmin + self.s_tmax = s_tmax + self.s_noise = s_noise + + def sampler_step(self, sigma, next_sigma, denoiser, x, cond, uc=None, gamma=0.0): + sigma_hat = sigma * (gamma + 1.0) + if gamma > 0: + eps = torch.randn_like(x) * self.s_noise + x = x + eps * append_dims(sigma_hat**2 - sigma**2, x.ndim) ** 0.5 + + denoised = self.denoise(x, denoiser, sigma_hat, cond, uc) + d = to_d(x, sigma_hat, denoised) + dt = append_dims(next_sigma - sigma_hat, x.ndim) + + euler_step = self.euler_step(x, d, dt) + x = self.possible_correction_step( + euler_step, x, d, dt, next_sigma, denoiser, cond, uc + ) + return x + + def __call__(self, denoiser, x, cond, uc=None, num_steps=None): + x, s_in, sigmas, num_sigmas, cond, uc = self.prepare_sampling_loop( + x, cond, uc, num_steps + ) + + for i in self.get_sigma_gen(num_sigmas): + gamma = ( + min(self.s_churn / (num_sigmas - 1), 2**0.5 - 1) + if self.s_tmin <= sigmas[i] <= self.s_tmax + else 0.0 + ) + x = self.sampler_step( + s_in * sigmas[i], + s_in * sigmas[i + 1], + denoiser, + x, + cond, + uc, + gamma, + ) + + return x + + +class AncestralSampler(SingleStepDiffusionSampler): + def __init__(self, eta=1.0, s_noise=1.0, *args, **kwargs): + super().__init__(*args, **kwargs) + + self.eta = eta + self.s_noise = s_noise + self.noise_sampler = lambda x: torch.randn_like(x) + + def ancestral_euler_step(self, x, denoised, sigma, sigma_down): + d = to_d(x, sigma, denoised) + dt = append_dims(sigma_down - sigma, x.ndim) + + return self.euler_step(x, d, dt) + + def ancestral_step(self, x, sigma, next_sigma, sigma_up): + x = torch.where( + append_dims(next_sigma, x.ndim) > 0.0, + x + self.noise_sampler(x) * self.s_noise * append_dims(sigma_up, x.ndim), + x, + ) + return x + + def __call__(self, denoiser, x, cond, uc=None, num_steps=None): + x, s_in, sigmas, num_sigmas, cond, uc = self.prepare_sampling_loop( + x, cond, uc, num_steps + ) + + for i in self.get_sigma_gen(num_sigmas): + x = self.sampler_step( + s_in * sigmas[i], + s_in * sigmas[i + 1], + denoiser, + x, + cond, + uc, + ) + + return x + + +class LinearMultistepSampler(BaseDiffusionSampler): + def __init__( + self, + order=4, + *args, + **kwargs, + ): + super().__init__(*args, **kwargs) + + self.order = order + + def __call__(self, denoiser, x, cond, uc=None, num_steps=None, **kwargs): + x, s_in, sigmas, num_sigmas, cond, uc = self.prepare_sampling_loop( + x, cond, uc, num_steps + ) + + ds = [] + sigmas_cpu = sigmas.detach().cpu().numpy() + for i in self.get_sigma_gen(num_sigmas): + sigma = s_in * sigmas[i] + denoised = denoiser( + *self.guider.prepare_inputs(x, sigma, cond, uc), **kwargs + ) + denoised = self.guider(denoised, sigma) + d = to_d(x, sigma, denoised) + ds.append(d) + if len(ds) > self.order: + ds.pop(0) + cur_order = min(i + 1, self.order) + coeffs = [ + linear_multistep_coeff(cur_order, sigmas_cpu, i, j) + for j in range(cur_order) + ] + x = x + sum(coeff * d for coeff, d in zip(coeffs, reversed(ds))) + + return x + + +class EulerEDMSampler(EDMSampler): + def possible_correction_step( + self, euler_step, x, d, dt, next_sigma, denoiser, cond, uc + ): + return euler_step + + +class HeunEDMSampler(EDMSampler): + def possible_correction_step( + self, euler_step, x, d, dt, next_sigma, denoiser, cond, uc + ): + if torch.sum(next_sigma) < 1e-14: + # Save a network evaluation if all noise levels are 0 + return euler_step + else: + denoised = self.denoise(euler_step, denoiser, next_sigma, cond, uc) + d_new = to_d(euler_step, next_sigma, denoised) + d_prime = (d + d_new) / 2.0 + + # apply correction if noise level is not 0 + x = torch.where( + append_dims(next_sigma, x.ndim) > 0.0, x + d_prime * dt, euler_step + ) + return x + + +class EulerAncestralSampler(AncestralSampler): + def sampler_step(self, sigma, next_sigma, denoiser, x, cond, uc): + sigma_down, sigma_up = get_ancestral_step(sigma, next_sigma, eta=self.eta) + denoised = self.denoise(x, denoiser, sigma, cond, uc) + x = self.ancestral_euler_step(x, denoised, sigma, sigma_down) + x = self.ancestral_step(x, sigma, next_sigma, sigma_up) + + return x + + +class DPMPP2SAncestralSampler(AncestralSampler): + def get_variables(self, sigma, sigma_down): + t, t_next = [to_neg_log_sigma(s) for s in (sigma, sigma_down)] + h = t_next - t + s = t + 0.5 * h + return h, s, t, t_next + + def get_mult(self, h, s, t, t_next): + mult1 = to_sigma(s) / to_sigma(t) + mult2 = (-0.5 * h).expm1() + mult3 = to_sigma(t_next) / to_sigma(t) + mult4 = (-h).expm1() + + return mult1, mult2, mult3, mult4 + + def sampler_step(self, sigma, next_sigma, denoiser, x, cond, uc=None, **kwargs): + sigma_down, sigma_up = get_ancestral_step(sigma, next_sigma, eta=self.eta) + denoised = self.denoise(x, denoiser, sigma, cond, uc) + x_euler = self.ancestral_euler_step(x, denoised, sigma, sigma_down) + + if torch.sum(sigma_down) < 1e-14: + # Save a network evaluation if all noise levels are 0 + x = x_euler + else: + h, s, t, t_next = self.get_variables(sigma, sigma_down) + mult = [ + append_dims(mult, x.ndim) for mult in self.get_mult(h, s, t, t_next) + ] + + x2 = mult[0] * x - mult[1] * denoised + denoised2 = self.denoise(x2, denoiser, to_sigma(s), cond, uc) + x_dpmpp2s = mult[2] * x - mult[3] * denoised2 + + # apply correction if noise level is not 0 + x = torch.where(append_dims(sigma_down, x.ndim) > 0.0, x_dpmpp2s, x_euler) + + x = self.ancestral_step(x, sigma, next_sigma, sigma_up) + return x + + +class DPMPP2MSampler(BaseDiffusionSampler): + def get_variables(self, sigma, next_sigma, previous_sigma=None): + t, t_next = [to_neg_log_sigma(s) for s in (sigma, next_sigma)] + h = t_next - t + + if previous_sigma is not None: + h_last = t - to_neg_log_sigma(previous_sigma) + r = h_last / h + return h, r, t, t_next + else: + return h, None, t, t_next + + def get_mult(self, h, r, t, t_next, previous_sigma): + mult1 = to_sigma(t_next) / to_sigma(t) + mult2 = (-h).expm1() + + if previous_sigma is not None: + mult3 = 1 + 1 / (2 * r) + mult4 = 1 / (2 * r) + return mult1, mult2, mult3, mult4 + else: + return mult1, mult2 + + def sampler_step( + self, + old_denoised, + previous_sigma, + sigma, + next_sigma, + denoiser, + x, + cond, + uc=None, + ): + denoised = self.denoise(x, denoiser, sigma, cond, uc) + + h, r, t, t_next = self.get_variables(sigma, next_sigma, previous_sigma) + mult = [ + append_dims(mult, x.ndim) + for mult in self.get_mult(h, r, t, t_next, previous_sigma) + ] + + x_standard = mult[0] * x - mult[1] * denoised + if old_denoised is None or torch.sum(next_sigma) < 1e-14: + # Save a network evaluation if all noise levels are 0 or on the first step + return x_standard, denoised + else: + denoised_d = mult[2] * denoised - mult[3] * old_denoised + x_advanced = mult[0] * x - mult[1] * denoised_d + + # apply correction if noise level is not 0 and not first step + x = torch.where( + append_dims(next_sigma, x.ndim) > 0.0, x_advanced, x_standard + ) + + return x, denoised + + def __call__(self, denoiser, x, cond, uc=None, num_steps=None, **kwargs): + x, s_in, sigmas, num_sigmas, cond, uc = self.prepare_sampling_loop( + x, cond, uc, num_steps + ) + + old_denoised = None + for i in self.get_sigma_gen(num_sigmas): + x, old_denoised = self.sampler_step( + old_denoised, + None if i == 0 else s_in * sigmas[i - 1], + s_in * sigmas[i], + s_in * sigmas[i + 1], + denoiser, + x, + cond, + uc=uc, + ) + + return x diff --git a/sgm/modules/diffusionmodules/sampling_utils.py b/sgm/modules/diffusionmodules/sampling_utils.py new file mode 100644 index 0000000..ce78527 --- /dev/null +++ b/sgm/modules/diffusionmodules/sampling_utils.py @@ -0,0 +1,43 @@ +import torch +from scipy import integrate + +from ...util import append_dims + + +def linear_multistep_coeff(order, t, i, j, epsrel=1e-4): + if order - 1 > i: + raise ValueError(f"Order {order} too high for step {i}") + + def fn(tau): + prod = 1.0 + for k in range(order): + if j == k: + continue + prod *= (tau - t[i - k]) / (t[i - j] - t[i - k]) + return prod + + return integrate.quad(fn, t[i], t[i + 1], epsrel=epsrel)[0] + + +def get_ancestral_step(sigma_from, sigma_to, eta=1.0): + if not eta: + return sigma_to, 0.0 + sigma_up = torch.minimum( + sigma_to, + eta + * (sigma_to**2 * (sigma_from**2 - sigma_to**2) / sigma_from**2) ** 0.5, + ) + sigma_down = (sigma_to**2 - sigma_up**2) ** 0.5 + return sigma_down, sigma_up + + +def to_d(x, sigma, denoised): + return (x - denoised) / append_dims(sigma, x.ndim) + + +def to_neg_log_sigma(sigma): + return sigma.log().neg() + + +def to_sigma(neg_log_sigma): + return neg_log_sigma.neg().exp() diff --git a/sgm/modules/diffusionmodules/sigma_sampling.py b/sgm/modules/diffusionmodules/sigma_sampling.py new file mode 100644 index 0000000..e327e85 --- /dev/null +++ b/sgm/modules/diffusionmodules/sigma_sampling.py @@ -0,0 +1,38 @@ +import torch +from typing import Optional, Union +from ...util import default, instantiate_from_config + + +class EDMSampling: + def __init__(self, p_mean=-1.2, p_std=1.2): + self.p_mean = p_mean + self.p_std = p_std + + def __call__(self, n_samples, rand=None): + log_sigma = self.p_mean + self.p_std * default(rand, torch.randn((n_samples,))) + return log_sigma.exp() + + +class DiscreteSampling: + def __init__(self, discretization_config, num_idx, do_append_zero=False, flip=True): + self.num_idx = num_idx + self.sigmas = instantiate_from_config(discretization_config)( + num_idx, do_append_zero=do_append_zero, flip=flip + ) + + def idx_to_sigma(self, idx): + return self.sigmas[idx] + + def __call__(self, n_samples, rand=None): + idx = default( + rand, + torch.randint(0, self.num_idx, (n_samples,)), + ) + return self.idx_to_sigma(idx) + + +class ZeroSampler: + def __call__( + self, n_samples: int, rand: Optional[torch.Tensor] = None + ) -> torch.Tensor: + return torch.zeros_like(default(rand, torch.randn((n_samples,)))) + 1.0e-5 diff --git a/sgm/modules/diffusionmodules/util.py b/sgm/modules/diffusionmodules/util.py new file mode 100644 index 0000000..1d06ce8 --- /dev/null +++ b/sgm/modules/diffusionmodules/util.py @@ -0,0 +1,399 @@ +""" +partially adopted from +https://github.com/openai/improved-diffusion/blob/main/improved_diffusion/gaussian_diffusion.py +and +https://github.com/lucidrains/denoising-diffusion-pytorch/blob/7706bdfc6f527f58d33f84b7b522e61e6e3164b3/denoising_diffusion_pytorch/denoising_diffusion_pytorch.py +and +https://github.com/openai/guided-diffusion/blob/0ba878e517b276c45d1195eb29f6f5f72659a05b/guided_diffusion/nn.py + +thanks! +""" + +import math +from typing import Optional + +import torch +import torch.nn as nn +from einops import rearrange, repeat + + +def get_alpha( + merge_strategy: str, + mix_factor: Optional[torch.Tensor], + image_only_indicator: torch.Tensor, + apply_sigmoid: bool = True, + is_attn: bool = False, +) -> torch.Tensor: + if merge_strategy == "fixed" or merge_strategy == "learned": + alpha = mix_factor + elif merge_strategy == "learned_with_images": + alpha = torch.where( + image_only_indicator.bool(), + torch.ones(1, 1, device=image_only_indicator.device), + rearrange(mix_factor, "... -> ... 1"), + ) + if is_attn: + alpha = rearrange(alpha, "b t -> (b t) 1 1") + else: + alpha = rearrange(alpha, "b t -> b 1 t 1 1") + elif merge_strategy == "fixed_with_images": + alpha = image_only_indicator + if is_attn: + alpha = rearrange(alpha, "b t -> (b t) 1 1") + else: + alpha = rearrange(alpha, "b t -> b 1 t 1 1") + else: + raise NotImplementedError + return torch.sigmoid(alpha) if apply_sigmoid else alpha + + +def make_beta_schedule( + schedule, + n_timestep, + linear_start=1e-4, + linear_end=2e-2, +): + if schedule == "linear": + betas = ( + torch.linspace( + linear_start**0.5, linear_end**0.5, n_timestep, dtype=torch.float64 + ) + ** 2 + ) + return betas.numpy() + + +def extract_into_tensor(a, t, x_shape): + b, *_ = t.shape + out = a.gather(-1, t) + return out.reshape(b, *((1,) * (len(x_shape) - 1))) + + +def mixed_checkpoint(func, inputs: dict, params, flag): + """ + Evaluate a function without caching intermediate activations, allowing for + reduced memory at the expense of extra compute in the backward pass. This differs from the original checkpoint function + borrowed from https://github.com/openai/guided-diffusion/blob/0ba878e517b276c45d1195eb29f6f5f72659a05b/guided_diffusion/nn.py in that + it also works with non-tensor inputs + :param func: the function to evaluate. + :param inputs: the argument dictionary to pass to `func`. + :param params: a sequence of parameters `func` depends on but does not + explicitly take as arguments. + :param flag: if False, disable gradient checkpointing. + """ + if flag: + tensor_keys = [key for key in inputs if isinstance(inputs[key], torch.Tensor)] + tensor_inputs = [ + inputs[key] for key in inputs if isinstance(inputs[key], torch.Tensor) + ] + non_tensor_keys = [ + key for key in inputs if not isinstance(inputs[key], torch.Tensor) + ] + non_tensor_inputs = [ + inputs[key] for key in inputs if not isinstance(inputs[key], torch.Tensor) + ] + args = tuple(tensor_inputs) + tuple(non_tensor_inputs) + tuple(params) + return MixedCheckpointFunction.apply( + func, + len(tensor_inputs), + len(non_tensor_inputs), + tensor_keys, + non_tensor_keys, + *args, + ) + else: + return func(**inputs) + + +class MixedCheckpointFunction(torch.autograd.Function): + @staticmethod + def forward( + ctx, + run_function, + length_tensors, + length_non_tensors, + tensor_keys, + non_tensor_keys, + *args, + ): + ctx.end_tensors = length_tensors + ctx.end_non_tensors = length_tensors + length_non_tensors + ctx.gpu_autocast_kwargs = { + "enabled": torch.is_autocast_enabled(), + "dtype": torch.get_autocast_gpu_dtype(), + "cache_enabled": torch.is_autocast_cache_enabled(), + } + assert ( + len(tensor_keys) == length_tensors + and len(non_tensor_keys) == length_non_tensors + ) + + ctx.input_tensors = { + key: val for (key, val) in zip(tensor_keys, list(args[: ctx.end_tensors])) + } + ctx.input_non_tensors = { + key: val + for (key, val) in zip( + non_tensor_keys, list(args[ctx.end_tensors : ctx.end_non_tensors]) + ) + } + ctx.run_function = run_function + ctx.input_params = list(args[ctx.end_non_tensors :]) + + with torch.no_grad(): + output_tensors = ctx.run_function( + **ctx.input_tensors, **ctx.input_non_tensors + ) + return output_tensors + + @staticmethod + def backward(ctx, *output_grads): + # additional_args = {key: ctx.input_tensors[key] for key in ctx.input_tensors if not isinstance(ctx.input_tensors[key],torch.Tensor)} + ctx.input_tensors = { + key: ctx.input_tensors[key].detach().requires_grad_(True) + for key in ctx.input_tensors + } + + with torch.enable_grad(), torch.cuda.amp.autocast(**ctx.gpu_autocast_kwargs): + # Fixes a bug where the first op in run_function modifies the + # Tensor storage in place, which is not allowed for detach()'d + # Tensors. + shallow_copies = { + key: ctx.input_tensors[key].view_as(ctx.input_tensors[key]) + for key in ctx.input_tensors + } + # shallow_copies.update(additional_args) + output_tensors = ctx.run_function(**shallow_copies, **ctx.input_non_tensors) + input_grads = torch.autograd.grad( + output_tensors, + list(ctx.input_tensors.values()) + ctx.input_params, + output_grads, + allow_unused=True, + ) + del ctx.input_tensors + del ctx.input_params + del output_tensors + return ( + (None, None, None, None, None) + + input_grads[: ctx.end_tensors] + + (None,) * (ctx.end_non_tensors - ctx.end_tensors) + + input_grads[ctx.end_tensors :] + ) + + +def checkpoint(func, inputs, params, flag): + """ + Evaluate a function without caching intermediate activations, allowing for + reduced memory at the expense of extra compute in the backward pass. + :param func: the function to evaluate. + :param inputs: the argument sequence to pass to `func`. + :param params: a sequence of parameters `func` depends on but does not + explicitly take as arguments. + :param flag: if False, disable gradient checkpointing. + """ + if flag: + args = tuple(inputs) + tuple(params) + return CheckpointFunction.apply(func, len(inputs), *args) + else: + return func(*inputs) + + +class CheckpointFunction(torch.autograd.Function): + @staticmethod + def forward(ctx, run_function, length, *args): + ctx.run_function = run_function + ctx.input_tensors = list(args[:length]) + ctx.input_params = list(args[length:]) + ctx.gpu_autocast_kwargs = { + "enabled": torch.is_autocast_enabled(), + "dtype": torch.get_autocast_gpu_dtype(), + "cache_enabled": torch.is_autocast_cache_enabled(), + } + with torch.no_grad(): + output_tensors = ctx.run_function(*ctx.input_tensors) + return output_tensors + + @staticmethod + def backward(ctx, *output_grads): + ctx.input_tensors = [x.detach().requires_grad_(True) for x in ctx.input_tensors] + with torch.enable_grad(), torch.cuda.amp.autocast(**ctx.gpu_autocast_kwargs): + # Fixes a bug where the first op in run_function modifies the + # Tensor storage in place, which is not allowed for detach()'d + # Tensors. + shallow_copies = [x.view_as(x) for x in ctx.input_tensors] + output_tensors = ctx.run_function(*shallow_copies) + input_grads = torch.autograd.grad( + output_tensors, + ctx.input_tensors + ctx.input_params, + output_grads, + allow_unused=True, + ) + del ctx.input_tensors + del ctx.input_params + del output_tensors + return (None, None) + input_grads + + +def timestep_embedding(timesteps, dim, max_period=10000, repeat_only=False): + """ + Create sinusoidal timestep embeddings. + :param timesteps: a 1-D Tensor of N indices, one per batch element. + These may be fractional. + :param dim: the dimension of the output. + :param max_period: controls the minimum frequency of the embeddings. + :return: an [N x dim] Tensor of positional embeddings. + """ + if not repeat_only: + half = dim // 2 + freqs = torch.exp( + -math.log(max_period) + * torch.arange(start=0, end=half, dtype=torch.float32) + / half + ).to(device=timesteps.device) + args = timesteps[:, None].float() * freqs[None] + embedding = torch.cat([torch.cos(args), torch.sin(args)], dim=-1) + if dim % 2: + embedding = torch.cat( + [embedding, torch.zeros_like(embedding[:, :1])], dim=-1 + ) + else: + embedding = repeat(timesteps, "b -> b d", d=dim) + return embedding + + +def zero_module(module): + """ + Zero out the parameters of a module and return it. + """ + for p in module.parameters(): + p.detach().zero_() + return module + + +def scale_module(module, scale): + """ + Scale the parameters of a module and return it. + """ + for p in module.parameters(): + p.detach().mul_(scale) + return module + + +def mean_flat(tensor): + """ + Take the mean over all non-batch dimensions. + """ + return tensor.mean(dim=list(range(1, len(tensor.shape)))) + + +def normalization(channels): + """ + Make a standard normalization layer. + :param channels: number of input channels. + :return: an nn.Module for normalization. + """ + return GroupNorm32(32, channels) + + +# PyTorch 1.7 has SiLU, but we support PyTorch 1.5. +class SiLU(nn.Module): + def forward(self, x): + return x * torch.sigmoid(x) + + +class GroupNorm32(nn.GroupNorm): + def forward(self, x): + return super().forward(x.float()).type(x.dtype) + + +def conv_nd(dims, *args, **kwargs): + """ + Create a 1D, 2D, or 3D convolution module. + """ + if dims == 1: + return nn.Conv1d(*args, **kwargs) + elif dims == 2: + return nn.Conv2d(*args, **kwargs) + elif dims == 3: + return nn.Conv3d(*args, **kwargs) + raise ValueError(f"unsupported dimensions: {dims}") + + +def linear(*args, **kwargs): + """ + Create a linear module. + """ + return nn.Linear(*args, **kwargs) + + +def avg_pool_nd(dims, *args, **kwargs): + """ + Create a 1D, 2D, or 3D average pooling module. + """ + if dims == 1: + return nn.AvgPool1d(*args, **kwargs) + elif dims == 2: + return nn.AvgPool2d(*args, **kwargs) + elif dims == 3: + return nn.AvgPool3d(*args, **kwargs) + raise ValueError(f"unsupported dimensions: {dims}") + + +class AlphaBlender(nn.Module): + strategies = ["learned", "fixed", "learned_with_images"] + + def __init__( + self, + alpha: float, + merge_strategy: str = "learned_with_images", + rearrange_pattern: str = "b t -> (b t) 1 1", + ): + super().__init__() + self.merge_strategy = merge_strategy + self.rearrange_pattern = rearrange_pattern + + assert ( + merge_strategy in self.strategies + ), f"merge_strategy needs to be in {self.strategies}" + + if self.merge_strategy == "fixed": + self.register_buffer("mix_factor", torch.Tensor([alpha])) + elif ( + self.merge_strategy == "learned" + or self.merge_strategy == "learned_with_images" + ): + self.register_parameter( + "mix_factor", torch.nn.Parameter(torch.Tensor([alpha])) + ) + else: + raise ValueError(f"unknown merge strategy {self.merge_strategy}") + + def get_alpha(self, image_only_indicator: torch.Tensor) -> torch.Tensor: + if self.merge_strategy == "fixed": + alpha = self.mix_factor + elif self.merge_strategy == "learned": + alpha = torch.sigmoid(self.mix_factor) + elif self.merge_strategy == "learned_with_images": + assert image_only_indicator is not None, "need image_only_indicator ..." + alpha = torch.where( + image_only_indicator.bool(), + torch.ones(1, 1, device=image_only_indicator.device), + rearrange(torch.sigmoid(self.mix_factor), "... -> ... 1"), + ) + alpha = rearrange(alpha, self.rearrange_pattern) + else: + raise NotImplementedError + return alpha + + def forward( + self, + x_spatial: torch.Tensor, + x_temporal: torch.Tensor, + image_only_indicator: Optional[torch.Tensor] = None, + ) -> torch.Tensor: + alpha = self.get_alpha(image_only_indicator) + x = ( + alpha.to(x_spatial.dtype) * x_spatial + + (1.0 - alpha).to(x_spatial.dtype) * x_temporal + ) + return x diff --git a/sgm/modules/diffusionmodules/video_model.py b/sgm/modules/diffusionmodules/video_model.py new file mode 100644 index 0000000..64b9844 --- /dev/null +++ b/sgm/modules/diffusionmodules/video_model.py @@ -0,0 +1,1241 @@ +from functools import partial +from typing import List, Optional, Union + +from einops import rearrange + +from ...modules.diffusionmodules.openaimodel import * +from ...modules.video_attention import SpatialVideoTransformer +from ...modules.spacetime_attention import ( + BasicTransformerTimeMixBlock, + PostHocSpatialTransformerWithTimeMixing, + PostHocSpatialTransformerWithTimeMixingAndMotion +) +from ...util import default +from .util import AlphaBlender # , LegacyAlphaBlenderWithBug, get_alpha + + +class VideoResBlock(ResBlock): + def __init__( + self, + channels: int, + emb_channels: int, + dropout: float, + video_kernel_size: Union[int, List[int]] = 3, + merge_strategy: str = "fixed", + merge_factor: float = 0.5, + out_channels: Optional[int] = None, + use_conv: bool = False, + use_scale_shift_norm: bool = False, + dims: int = 2, + use_checkpoint: bool = False, + up: bool = False, + down: bool = False, + ): + super().__init__( + channels, + emb_channels, + dropout, + out_channels=out_channels, + use_conv=use_conv, + use_scale_shift_norm=use_scale_shift_norm, + dims=dims, + use_checkpoint=use_checkpoint, + up=up, + down=down, + ) + + self.time_stack = ResBlock( + default(out_channels, channels), + emb_channels, + dropout=dropout, + dims=3, + out_channels=default(out_channels, channels), + use_scale_shift_norm=False, + use_conv=False, + up=False, + down=False, + kernel_size=video_kernel_size, + use_checkpoint=use_checkpoint, + exchange_temb_dims=True, + ) + self.time_mixer = AlphaBlender( + alpha=merge_factor, + merge_strategy=merge_strategy, + rearrange_pattern="b t -> b 1 t 1 1", + ) + + def forward( + self, + x: th.Tensor, + emb: th.Tensor, + num_video_frames: int, + image_only_indicator: Optional[th.Tensor] = None, + ) -> th.Tensor: + x = super().forward(x, emb) + + x_mix = rearrange(x, "(b t) c h w -> b c t h w", t=num_video_frames) + x = rearrange(x, "(b t) c h w -> b c t h w", t=num_video_frames) + + x = self.time_stack( + x, rearrange(emb, "(b t) ... -> b t ...", t=num_video_frames) + ) + x = self.time_mixer( + x_spatial=x_mix, x_temporal=x, image_only_indicator=image_only_indicator + ) + x = rearrange(x, "b c t h w -> (b t) c h w") + return x + + +class VideoUNet(nn.Module): + def __init__( + self, + in_channels: int, + model_channels: int, + out_channels: int, + num_res_blocks: int, + attention_resolutions: int, + dropout: float = 0.0, + channel_mult: List[int] = (1, 2, 4, 8), + conv_resample: bool = True, + dims: int = 2, + num_classes: Optional[int] = None, + use_checkpoint: bool = False, + num_heads: int = -1, + num_head_channels: int = -1, + num_heads_upsample: int = -1, + use_scale_shift_norm: bool = False, + resblock_updown: bool = False, + transformer_depth: Union[List[int], int] = 1, + transformer_depth_middle: Optional[int] = None, + context_dim: Optional[int] = None, + time_downup: bool = False, + time_context_dim: Optional[int] = None, + extra_ff_mix_layer: bool = False, + use_spatial_context: bool = False, + merge_strategy: str = "fixed", + merge_factor: float = 0.5, + spatial_transformer_attn_type: str = "softmax", + video_kernel_size: Union[int, List[int]] = 3, + use_linear_in_transformer: bool = False, + adm_in_channels: Optional[int] = None, + disable_temporal_crossattention: bool = False, + max_ddpm_temb_period: int = 10000, + ): + super().__init__() + assert context_dim is not None + + if num_heads_upsample == -1: + num_heads_upsample = num_heads + + if num_heads == -1: + assert num_head_channels != -1 + + if num_head_channels == -1: + assert num_heads != -1 + + self.in_channels = in_channels + self.model_channels = model_channels + self.out_channels = out_channels + if isinstance(transformer_depth, int): + transformer_depth = len(channel_mult) * [transformer_depth] + transformer_depth_middle = default( + transformer_depth_middle, transformer_depth[-1] + ) + + self.num_res_blocks = num_res_blocks + self.attention_resolutions = attention_resolutions + self.dropout = dropout + self.channel_mult = channel_mult + self.conv_resample = conv_resample + self.num_classes = num_classes + self.use_checkpoint = use_checkpoint + self.num_heads = num_heads + self.num_head_channels = num_head_channels + self.num_heads_upsample = num_heads_upsample + + time_embed_dim = model_channels * 4 + self.time_embed = nn.Sequential( + linear(model_channels, time_embed_dim), + nn.SiLU(), + linear(time_embed_dim, time_embed_dim), + ) + + if self.num_classes is not None: + if isinstance(self.num_classes, int): + self.label_emb = nn.Embedding(num_classes, time_embed_dim) + elif self.num_classes == "continuous": + print("setting up linear c_adm embedding layer") + self.label_emb = nn.Linear(1, time_embed_dim) + elif self.num_classes == "timestep": + self.label_emb = nn.Sequential( + Timestep(model_channels), + nn.Sequential( + linear(model_channels, time_embed_dim), + nn.SiLU(), + linear(time_embed_dim, time_embed_dim), + ), + ) + + elif self.num_classes == "sequential": + assert adm_in_channels is not None + self.label_emb = nn.Sequential( + nn.Sequential( + linear(adm_in_channels, time_embed_dim), + nn.SiLU(), + linear(time_embed_dim, time_embed_dim), + ) + ) + else: + raise ValueError() + + self.input_blocks = nn.ModuleList( + [ + TimestepEmbedSequential( + conv_nd(dims, in_channels, model_channels, 3, padding=1) + ) + ] + ) + self._feature_size = model_channels + input_block_chans = [model_channels] + ch = model_channels + ds = 1 + + def get_attention_layer( + ch, + num_heads, + dim_head, + depth=1, + context_dim=None, + use_checkpoint=False, + disabled_sa=False, + ): + return SpatialVideoTransformer( + ch, + num_heads, + dim_head, + depth=depth, + context_dim=context_dim, + time_context_dim=time_context_dim, + dropout=dropout, + ff_in=extra_ff_mix_layer, + use_spatial_context=use_spatial_context, + merge_strategy=merge_strategy, + merge_factor=merge_factor, + checkpoint=use_checkpoint, + use_linear=use_linear_in_transformer, + attn_mode=spatial_transformer_attn_type, + disable_self_attn=disabled_sa, + disable_temporal_crossattention=disable_temporal_crossattention, + max_time_embed_period=max_ddpm_temb_period, + ) + + def get_resblock( + merge_factor, + merge_strategy, + video_kernel_size, + ch, + time_embed_dim, + dropout, + out_ch, + dims, + use_checkpoint, + use_scale_shift_norm, + down=False, + up=False, + ): + return VideoResBlock( + merge_factor=merge_factor, + merge_strategy=merge_strategy, + video_kernel_size=video_kernel_size, + channels=ch, + emb_channels=time_embed_dim, + dropout=dropout, + out_channels=out_ch, + dims=dims, + use_checkpoint=use_checkpoint, + use_scale_shift_norm=use_scale_shift_norm, + down=down, + up=up, + ) + + for level, mult in enumerate(channel_mult): + for _ in range(num_res_blocks): + layers = [ + get_resblock( + merge_factor=merge_factor, + merge_strategy=merge_strategy, + video_kernel_size=video_kernel_size, + ch=ch, + time_embed_dim=time_embed_dim, + dropout=dropout, + out_ch=mult * model_channels, + dims=dims, + use_checkpoint=use_checkpoint, + use_scale_shift_norm=use_scale_shift_norm, + ) + ] + ch = mult * model_channels + if ds in attention_resolutions: + if num_head_channels == -1: + dim_head = ch // num_heads + else: + num_heads = ch // num_head_channels + dim_head = num_head_channels + + layers.append( + get_attention_layer( + ch, + num_heads, + dim_head, + depth=transformer_depth[level], + context_dim=context_dim, + use_checkpoint=use_checkpoint, + disabled_sa=False, + ) + ) + self.input_blocks.append(TimestepEmbedSequential(*layers)) + self._feature_size += ch + input_block_chans.append(ch) + if level != len(channel_mult) - 1: + ds *= 2 + out_ch = ch + self.input_blocks.append( + TimestepEmbedSequential( + get_resblock( + merge_factor=merge_factor, + merge_strategy=merge_strategy, + video_kernel_size=video_kernel_size, + ch=ch, + time_embed_dim=time_embed_dim, + dropout=dropout, + out_ch=out_ch, + dims=dims, + use_checkpoint=use_checkpoint, + use_scale_shift_norm=use_scale_shift_norm, + down=True, + ) + if resblock_updown + else Downsample( + ch, + conv_resample, + dims=dims, + out_channels=out_ch, + third_down=time_downup, + ) + ) + ) + ch = out_ch + input_block_chans.append(ch) + + self._feature_size += ch + + if num_head_channels == -1: + dim_head = ch // num_heads + else: + num_heads = ch // num_head_channels + dim_head = num_head_channels + + self.middle_block = TimestepEmbedSequential( + get_resblock( + merge_factor=merge_factor, + merge_strategy=merge_strategy, + video_kernel_size=video_kernel_size, + ch=ch, + time_embed_dim=time_embed_dim, + out_ch=None, + dropout=dropout, + dims=dims, + use_checkpoint=use_checkpoint, + use_scale_shift_norm=use_scale_shift_norm, + ), + get_attention_layer( + ch, + num_heads, + dim_head, + depth=transformer_depth_middle, + context_dim=context_dim, + use_checkpoint=use_checkpoint, + ), + get_resblock( + merge_factor=merge_factor, + merge_strategy=merge_strategy, + video_kernel_size=video_kernel_size, + ch=ch, + out_ch=None, + time_embed_dim=time_embed_dim, + dropout=dropout, + dims=dims, + use_checkpoint=use_checkpoint, + use_scale_shift_norm=use_scale_shift_norm, + ), + ) + self._feature_size += ch + + self.output_blocks = nn.ModuleList([]) + for level, mult in list(enumerate(channel_mult))[::-1]: + for i in range(num_res_blocks + 1): + ich = input_block_chans.pop() + layers = [ + get_resblock( + merge_factor=merge_factor, + merge_strategy=merge_strategy, + video_kernel_size=video_kernel_size, + ch=ch + ich, + time_embed_dim=time_embed_dim, + dropout=dropout, + out_ch=model_channels * mult, + dims=dims, + use_checkpoint=use_checkpoint, + use_scale_shift_norm=use_scale_shift_norm, + ) + ] + ch = model_channels * mult + if ds in attention_resolutions: + if num_head_channels == -1: + dim_head = ch // num_heads + else: + num_heads = ch // num_head_channels + dim_head = num_head_channels + + layers.append( + get_attention_layer( + ch, + num_heads, + dim_head, + depth=transformer_depth[level], + context_dim=context_dim, + use_checkpoint=use_checkpoint, + disabled_sa=False, + ) + ) + if level and i == num_res_blocks: + out_ch = ch + ds //= 2 + layers.append( + get_resblock( + merge_factor=merge_factor, + merge_strategy=merge_strategy, + video_kernel_size=video_kernel_size, + ch=ch, + time_embed_dim=time_embed_dim, + dropout=dropout, + out_ch=out_ch, + dims=dims, + use_checkpoint=use_checkpoint, + use_scale_shift_norm=use_scale_shift_norm, + up=True, + ) + if resblock_updown + else Upsample( + ch, + conv_resample, + dims=dims, + out_channels=out_ch, + third_up=time_downup, + ) + ) + + self.output_blocks.append(TimestepEmbedSequential(*layers)) + self._feature_size += ch + + self.out = nn.Sequential( + normalization(ch), + nn.SiLU(), + zero_module(conv_nd(dims, model_channels, out_channels, 3, padding=1)), + ) + + def forward( + self, + x: th.Tensor, + timesteps: th.Tensor, + context: Optional[th.Tensor] = None, + y: Optional[th.Tensor] = None, + time_context: Optional[th.Tensor] = None, + num_video_frames: Optional[int] = None, + image_only_indicator: Optional[th.Tensor] = None, + ): + assert (y is not None) == ( + self.num_classes is not None + ), "must specify y if and only if the model is class-conditional -> no, relax this TODO" + hs = [] + t_emb = timestep_embedding(timesteps, self.model_channels, repeat_only=False) + emb = self.time_embed(t_emb) + + if self.num_classes is not None: + assert y.shape[0] == x.shape[0] + emb = emb + self.label_emb(y) + + h = x + for module in self.input_blocks: + h = module( + h, + emb, + context=context, + image_only_indicator=image_only_indicator, + time_context=time_context, + num_video_frames=num_video_frames, + ) + hs.append(h) + h = self.middle_block( + h, + emb, + context=context, + image_only_indicator=image_only_indicator, + time_context=time_context, + num_video_frames=num_video_frames, + ) + for module in self.output_blocks: + h = th.cat([h, hs.pop()], dim=1) + h = module( + h, + emb, + context=context, + image_only_indicator=image_only_indicator, + time_context=time_context, + num_video_frames=num_video_frames, + ) + h = h.type(x.dtype) + return self.out(h) + + +class PostHocAttentionBlockWithTimeMixing(AttentionBlock): + def __init__( + self, + in_channels: int, + n_heads: int, + d_head: int, + use_checkpoint: bool = False, + use_new_attention_order: bool = False, + dropout: float = 0.0, + use_spatial_context: bool = False, + merge_strategy: bool = "fixed", + merge_factor: float = 0.5, + apply_sigmoid_to_merge: bool = True, + ff_in: bool = False, + attn_mode: str = "softmax", + disable_temporal_crossattention: bool = False, + ): + super().__init__( + in_channels, + n_heads, + d_head, + use_checkpoint=use_checkpoint, + use_new_attention_order=use_new_attention_order, + ) + inner_dim = n_heads * d_head + + self.time_mix_blocks = nn.ModuleList( + [ + BasicTransformerTimeMixBlock( + inner_dim, + n_heads, + d_head, + dropout=dropout, + checkpoint=use_checkpoint, + ff_in=ff_in, + attn_mode=attn_mode, + disable_temporal_crossattention=disable_temporal_crossattention, + ) + ] + ) + self.in_channels = in_channels + + time_embed_dim = self.in_channels * 4 + self.time_mix_time_embed = nn.Sequential( + linear(self.in_channels, time_embed_dim), + nn.SiLU(), + linear(time_embed_dim, self.in_channels), + ) + + self.use_spatial_context = use_spatial_context + + if merge_strategy == "fixed": + self.register_buffer("mix_factor", th.Tensor([merge_factor])) + elif merge_strategy == "learned" or merge_strategy == "learned_with_images": + self.register_parameter( + "mix_factor", th.nn.Parameter(th.Tensor([merge_factor])) + ) + elif merge_strategy == "fixed_with_images": + self.mix_factor = None + else: + raise ValueError(f"unknown merge strategy {merge_strategy}") + + self.get_alpha_fn = functools.partial( + get_alpha, + merge_strategy, + self.mix_factor, + apply_sigmoid=apply_sigmoid_to_merge, + ) + + def forward( + self, + x: th.Tensor, + context: Optional[th.Tensor] = None, + # cam: Optional[th.Tensor] = None, + time_context: Optional[th.Tensor] = None, + timesteps: Optional[int] = None, + image_only_indicator: Optional[th.Tensor] = None, + conv_view: Optional[th.Tensor] = None, + conv_motion: Optional[th.Tensor] = None, + ): + if time_context is not None: + raise NotImplementedError + + _, _, h, w = x.shape + if exists(context): + context = rearrange(context, "b t ... -> (b t) ...") + if self.use_spatial_context: + time_context = repeat(context[:, 0], "b ... -> (b n) ...", n=h * w) + + x = super().forward( + x, + ) + + x = rearrange(x, "b c h w -> b (h w) c") + x_mix = x + + num_frames = th.arange(timesteps, device=x.device) + num_frames = repeat(num_frames, "t -> b t", b=x.shape[0] // timesteps) + num_frames = rearrange(num_frames, "b t -> (b t)") + t_emb = timestep_embedding(num_frames, self.in_channels, repeat_only=False) + emb = self.time_mix_time_embed(t_emb) + emb = emb[:, None, :] + x_mix = x_mix + emb + + x_mix = self.time_mix_blocks[0]( + x_mix, context=time_context, timesteps=timesteps + ) + + alpha = self.get_alpha_fn(image_only_indicator=image_only_indicator) + x = alpha * x + (1.0 - alpha) * x_mix + x = rearrange(x, "b (h w) c -> b c h w", h=h, w=w) + return x + + +class PostHocResBlockWithTime(ResBlock): + def __init__( + self, + channels: int, + emb_channels: int, + dropout: float, + time_kernel_size: Union[int, List[int]] = 3, + merge_strategy: bool = "fixed", + merge_factor: float = 0.5, + apply_sigmoid_to_merge: bool = True, + out_channels: Optional[int] = None, + use_conv: bool = False, + use_scale_shift_norm: bool = False, + dims: int = 2, + use_checkpoint: bool = False, + up: bool = False, + down: bool = False, + time_mix_legacy: bool = True, + replicate_bug: bool = False, + ): + super().__init__( + channels, + emb_channels, + dropout, + out_channels=out_channels, + use_conv=use_conv, + use_scale_shift_norm=use_scale_shift_norm, + dims=dims, + use_checkpoint=use_checkpoint, + up=up, + down=down, + ) + + self.time_mix_blocks = ResBlock( + default(out_channels, channels), + emb_channels, + dropout=dropout, + dims=3, + out_channels=default(out_channels, channels), + use_scale_shift_norm=False, + use_conv=False, + up=False, + down=False, + kernel_size=time_kernel_size, + use_checkpoint=use_checkpoint, + exchange_temb_dims=True, + ) + self.time_mix_legacy = time_mix_legacy + if self.time_mix_legacy: + if merge_strategy == "fixed": + self.register_buffer("mix_factor", th.Tensor([merge_factor])) + elif merge_strategy == "learned" or merge_strategy == "learned_with_images": + self.register_parameter( + "mix_factor", th.nn.Parameter(th.Tensor([merge_factor])) + ) + elif merge_strategy == "fixed_with_images": + self.mix_factor = None + else: + raise ValueError(f"unknown merge strategy {merge_strategy}") + + self.get_alpha_fn = functools.partial( + get_alpha, + merge_strategy, + self.mix_factor, + apply_sigmoid=apply_sigmoid_to_merge, + ) + else: + if False: # replicate_bug: + logpy.warning( + "*****************************************************************************************\n" + "GRAVE WARNING: YOU'RE USING THE BUGGY LEGACY ALPHABLENDER!!! ARE YOU SURE YOU WANT THIS?!\n" + "*****************************************************************************************" + ) + self.time_mixer = LegacyAlphaBlenderWithBug( + alpha=merge_factor, + merge_strategy=merge_strategy, + rearrange_pattern="b t -> b 1 t 1 1", + ) + else: + self.time_mixer = AlphaBlender( + alpha=merge_factor, + merge_strategy=merge_strategy, + rearrange_pattern="b t -> b 1 t 1 1", + ) + + def forward( + self, + x: th.Tensor, + emb: th.Tensor, + num_video_frames: int, + image_only_indicator: Optional[th.Tensor] = None, + cond_view: Optional[th.Tensor] = None, + cond_motion: Optional[th.Tensor] = None, + ) -> th.Tensor: + x = super().forward(x, emb) + + x_mix = rearrange(x, "(b t) c h w -> b c t h w", t=num_video_frames) + x = rearrange(x, "(b t) c h w -> b c t h w", t=num_video_frames) + + x = self.time_mix_blocks( + x, rearrange(emb, "(b t) ... -> b t ...", t=num_video_frames) + ) + + if self.time_mix_legacy: + alpha = self.get_alpha_fn(image_only_indicator=image_only_indicator) + x = alpha.to(x.dtype) * x + (1.0 - alpha).to(x.dtype) * x_mix + else: + x = self.time_mixer( + x_spatial=x_mix, x_temporal=x, image_only_indicator=image_only_indicator + ) + x = rearrange(x, "b c t h w -> (b t) c h w") + return x + + +class SpatialUNetModelWithTime(nn.Module): + def __init__( + self, + in_channels: int, + model_channels: int, + out_channels: int, + num_res_blocks: int, + attention_resolutions: int, + dropout: float = 0.0, + channel_mult: List[int] = (1, 2, 4, 8), + conv_resample: bool = True, + dims: int = 2, + num_classes: Optional[int] = None, + use_checkpoint: bool = False, + num_heads: int = -1, + num_head_channels: int = -1, + num_heads_upsample: int = -1, + use_scale_shift_norm: bool = False, + resblock_updown: bool = False, + use_new_attention_order: bool = False, + use_spatial_transformer: bool = False, + transformer_depth: Union[List[int], int] = 1, + transformer_depth_middle: Optional[int] = None, + context_dim: Optional[int] = None, + time_downup: bool = False, + time_context_dim: Optional[int] = None, + extra_ff_mix_layer: bool = False, + use_spatial_context: bool = False, + time_block_merge_strategy: str = "fixed", + time_block_merge_factor: float = 0.5, + spatial_transformer_attn_type: str = "softmax", + time_kernel_size: Union[int, List[int]] = 3, + use_linear_in_transformer: bool = False, + legacy: bool = True, + adm_in_channels: Optional[int] = None, + use_temporal_resblock: bool = True, + disable_temporal_crossattention: bool = False, + time_mix_legacy: bool = True, + max_ddpm_temb_period: int = 10000, + replicate_time_mix_bug: bool = False, + use_motion_attention: bool = False, + ): + super().__init__() + + if use_spatial_transformer: + assert context_dim is not None + + if context_dim is not None: + assert use_spatial_transformer + + if num_heads_upsample == -1: + num_heads_upsample = num_heads + + if num_heads == -1: + assert num_head_channels != -1 + + if num_head_channels == -1: + assert num_heads != -1 + + self.in_channels = in_channels + self.model_channels = model_channels + self.out_channels = out_channels + if isinstance(transformer_depth, int): + transformer_depth = len(channel_mult) * [transformer_depth] + transformer_depth_middle = default( + transformer_depth_middle, transformer_depth[-1] + ) + + self.num_res_blocks = num_res_blocks + self.attention_resolutions = attention_resolutions + self.dropout = dropout + self.channel_mult = channel_mult + self.conv_resample = conv_resample + self.num_classes = num_classes + self.use_checkpoint = use_checkpoint + self.num_heads = num_heads + self.num_head_channels = num_head_channels + self.num_heads_upsample = num_heads_upsample + self.use_temporal_resblocks = use_temporal_resblock + + time_embed_dim = model_channels * 4 + self.time_embed = nn.Sequential( + linear(model_channels, time_embed_dim), + nn.SiLU(), + linear(time_embed_dim, time_embed_dim), + ) + + if self.num_classes is not None: + if isinstance(self.num_classes, int): + self.label_emb = nn.Embedding(num_classes, time_embed_dim) + elif self.num_classes == "continuous": + print("setting up linear c_adm embedding layer") + self.label_emb = nn.Linear(1, time_embed_dim) + elif self.num_classes == "timestep": + self.label_emb = nn.Sequential( + Timestep(model_channels), + nn.Sequential( + linear(model_channels, time_embed_dim), + nn.SiLU(), + linear(time_embed_dim, time_embed_dim), + ), + ) + + elif self.num_classes == "sequential": + assert adm_in_channels is not None + self.label_emb = nn.Sequential( + nn.Sequential( + linear(adm_in_channels, time_embed_dim), + nn.SiLU(), + linear(time_embed_dim, time_embed_dim), + ) + ) + else: + raise ValueError() + + self.input_blocks = nn.ModuleList( + [ + TimestepEmbedSequential( + conv_nd(dims, in_channels, model_channels, 3, padding=1) + ) + ] + ) + self._feature_size = model_channels + input_block_chans = [model_channels] + ch = model_channels + ds = 1 + + def get_attention_layer( + ch, + num_heads, + dim_head, + depth=1, + context_dim=None, + use_checkpoint=False, + disabled_sa=False, + ): + if not use_spatial_transformer: + return PostHocAttentionBlockWithTimeMixing( + ch, + num_heads, + dim_head, + use_checkpoint=use_checkpoint, + use_new_attention_order=use_new_attention_order, + dropout=dropout, + ff_in=extra_ff_mix_layer, + use_spatial_context=use_spatial_context, + merge_strategy=time_block_merge_strategy, + merge_factor=time_block_merge_factor, + attn_mode=spatial_transformer_attn_type, + disable_temporal_crossattention=disable_temporal_crossattention, + ) + + elif use_motion_attention: + return PostHocSpatialTransformerWithTimeMixingAndMotion( + ch, + num_heads, + dim_head, + depth=depth, + context_dim=context_dim, + time_context_dim=time_context_dim, + dropout=dropout, + ff_in=extra_ff_mix_layer, + use_spatial_context=use_spatial_context, + merge_strategy=time_block_merge_strategy, + merge_factor=time_block_merge_factor, + checkpoint=use_checkpoint, + use_linear=use_linear_in_transformer, + attn_mode=spatial_transformer_attn_type, + disable_self_attn=disabled_sa, + disable_temporal_crossattention=disable_temporal_crossattention, + time_mix_legacy=time_mix_legacy, + max_time_embed_period=max_ddpm_temb_period, + ) + + else: + return PostHocSpatialTransformerWithTimeMixing( + ch, + num_heads, + dim_head, + depth=depth, + context_dim=context_dim, + time_context_dim=time_context_dim, + dropout=dropout, + ff_in=extra_ff_mix_layer, + use_spatial_context=use_spatial_context, + merge_strategy=time_block_merge_strategy, + merge_factor=time_block_merge_factor, + checkpoint=use_checkpoint, + use_linear=use_linear_in_transformer, + attn_mode=spatial_transformer_attn_type, + disable_self_attn=disabled_sa, + disable_temporal_crossattention=disable_temporal_crossattention, + time_mix_legacy=time_mix_legacy, + max_time_embed_period=max_ddpm_temb_period, + ) + + def get_resblock( + time_block_merge_factor, + time_block_merge_strategy, + time_kernel_size, + ch, + time_embed_dim, + dropout, + out_ch, + dims, + use_checkpoint, + use_scale_shift_norm, + down=False, + up=False, + ): + if self.use_temporal_resblocks: + return PostHocResBlockWithTime( + merge_factor=time_block_merge_factor, + merge_strategy=time_block_merge_strategy, + time_kernel_size=time_kernel_size, + channels=ch, + emb_channels=time_embed_dim, + dropout=dropout, + out_channels=out_ch, + dims=dims, + use_checkpoint=use_checkpoint, + use_scale_shift_norm=use_scale_shift_norm, + down=down, + up=up, + time_mix_legacy=time_mix_legacy, + replicate_bug=replicate_time_mix_bug, + ) + else: + return ResBlock( + channels=ch, + emb_channels=time_embed_dim, + dropout=dropout, + out_channels=out_ch, + use_checkpoint=use_checkpoint, + dims=dims, + use_scale_shift_norm=use_scale_shift_norm, + down=down, + up=up, + ) + + for level, mult in enumerate(channel_mult): + for _ in range(num_res_blocks): + layers = [ + get_resblock( + time_block_merge_factor=time_block_merge_factor, + time_block_merge_strategy=time_block_merge_strategy, + time_kernel_size=time_kernel_size, + ch=ch, + time_embed_dim=time_embed_dim, + dropout=dropout, + out_ch=mult * model_channels, + dims=dims, + use_checkpoint=use_checkpoint, + use_scale_shift_norm=use_scale_shift_norm, + ) + ] + ch = mult * model_channels + if ds in attention_resolutions: + if num_head_channels == -1: + dim_head = ch // num_heads + else: + num_heads = ch // num_head_channels + dim_head = num_head_channels + if legacy: + dim_head = ( + ch // num_heads + if use_spatial_transformer + else num_head_channels + ) + + layers.append( + get_attention_layer( + ch, + num_heads, + dim_head, + depth=transformer_depth[level], + context_dim=context_dim, + use_checkpoint=use_checkpoint, + disabled_sa=False, + ) + ) + self.input_blocks.append(TimestepEmbedSequential(*layers)) + self._feature_size += ch + input_block_chans.append(ch) + if level != len(channel_mult) - 1: + ds *= 2 + out_ch = ch + self.input_blocks.append( + TimestepEmbedSequential( + get_resblock( + time_block_merge_factor=time_block_merge_factor, + time_block_merge_strategy=time_block_merge_strategy, + time_kernel_size=time_kernel_size, + ch=ch, + time_embed_dim=time_embed_dim, + dropout=dropout, + out_ch=out_ch, + dims=dims, + use_checkpoint=use_checkpoint, + use_scale_shift_norm=use_scale_shift_norm, + down=True, + ) + if resblock_updown + else Downsample( + ch, + conv_resample, + dims=dims, + out_channels=out_ch, + third_down=time_downup, + ) + ) + ) + ch = out_ch + input_block_chans.append(ch) + + self._feature_size += ch + + if num_head_channels == -1: + dim_head = ch // num_heads + else: + num_heads = ch // num_head_channels + dim_head = num_head_channels + if legacy: + # num_heads = 1 + dim_head = ch // num_heads if use_spatial_transformer else num_head_channels + + self.middle_block = TimestepEmbedSequential( + get_resblock( + time_block_merge_factor=time_block_merge_factor, + time_block_merge_strategy=time_block_merge_strategy, + time_kernel_size=time_kernel_size, + ch=ch, + time_embed_dim=time_embed_dim, + out_ch=None, + dropout=dropout, + dims=dims, + use_checkpoint=use_checkpoint, + use_scale_shift_norm=use_scale_shift_norm, + ), + get_attention_layer( + ch, + num_heads, + dim_head, + depth=transformer_depth_middle, + context_dim=context_dim, + use_checkpoint=use_checkpoint, + ), + get_resblock( + time_block_merge_factor=time_block_merge_factor, + time_block_merge_strategy=time_block_merge_strategy, + time_kernel_size=time_kernel_size, + ch=ch, + out_ch=None, + time_embed_dim=time_embed_dim, + dropout=dropout, + dims=dims, + use_checkpoint=use_checkpoint, + use_scale_shift_norm=use_scale_shift_norm, + ), + ) + self._feature_size += ch + + self.output_blocks = nn.ModuleList([]) + for level, mult in list(enumerate(channel_mult))[::-1]: + for i in range(num_res_blocks + 1): + ich = input_block_chans.pop() + layers = [ + get_resblock( + time_block_merge_factor=time_block_merge_factor, + time_block_merge_strategy=time_block_merge_strategy, + time_kernel_size=time_kernel_size, + ch=ch + ich, + time_embed_dim=time_embed_dim, + dropout=dropout, + out_ch=model_channels * mult, + dims=dims, + use_checkpoint=use_checkpoint, + use_scale_shift_norm=use_scale_shift_norm, + ) + ] + ch = model_channels * mult + if ds in attention_resolutions: + if num_head_channels == -1: + dim_head = ch // num_heads + else: + num_heads = ch // num_head_channels + dim_head = num_head_channels + if legacy: + dim_head = ( + ch // num_heads + if use_spatial_transformer + else num_head_channels + ) + + layers.append( + get_attention_layer( + ch, + num_heads, + dim_head, + depth=transformer_depth[level], + context_dim=context_dim, + use_checkpoint=use_checkpoint, + disabled_sa=False, + ) + ) + if level and i == num_res_blocks: + out_ch = ch + ds //= 2 + layers.append( + get_resblock( + time_block_merge_factor=time_block_merge_factor, + time_block_merge_strategy=time_block_merge_strategy, + time_kernel_size=time_kernel_size, + ch=ch, + time_embed_dim=time_embed_dim, + dropout=dropout, + out_ch=out_ch, + dims=dims, + use_checkpoint=use_checkpoint, + use_scale_shift_norm=use_scale_shift_norm, + up=True, + ) + if resblock_updown + else Upsample( + ch, + conv_resample, + dims=dims, + out_channels=out_ch, + third_up=time_downup, + ) + ) + + self.output_blocks.append(TimestepEmbedSequential(*layers)) + self._feature_size += ch + + self.out = nn.Sequential( + normalization(ch), + nn.SiLU(), + zero_module(conv_nd(dims, model_channels, out_channels, 3, padding=1)), + ) + + def forward( + self, + x: th.Tensor, + timesteps: th.Tensor, + context: Optional[th.Tensor] = None, + y: Optional[th.Tensor] = None, + # cam: Optional[th.Tensor] = None, + time_context: Optional[th.Tensor] = None, + num_video_frames: Optional[int] = None, + image_only_indicator: Optional[th.Tensor] = None, + cond_view: Optional[th.Tensor] = None, + cond_motion: Optional[th.Tensor] = None, + time_step: Optional[int] = None, + ): + assert (y is not None) == ( + self.num_classes is not None + ), "must specify y if and only if the model is class-conditional -> no, relax this TODO" + hs = [] + t_emb = timestep_embedding(timesteps, self.model_channels, repeat_only=False) # 21 x 320 + emb = self.time_embed(t_emb) # 21 x 1280 + time = str(timesteps[0].data.cpu().numpy()) + + if self.num_classes is not None: + assert y.shape[0] == x.shape[0] + emb = emb + self.label_emb(y) # 21 x 1280 + + h = x # 21 x 8 x 64 x 64 + for i, module in enumerate(self.input_blocks): + h = module( + h, + emb, + context=context, + # cam=cam, + image_only_indicator=image_only_indicator, + cond_view=cond_view, + cond_motion=cond_motion, + time_context=time_context, + num_video_frames=num_video_frames, + time_step=time_step, + name='encoder_{}_{}'.format(time, i) + ) + hs.append(h) + h = self.middle_block( + h, + emb, + context=context, + # cam=cam, + image_only_indicator=image_only_indicator, + cond_view=cond_view, + cond_motion=cond_motion, + time_context=time_context, + num_video_frames=num_video_frames, + time_step=time_step, + name='middle_{}_0'.format(time, i) + ) + for i, module in enumerate(self.output_blocks): + h = th.cat([h, hs.pop()], dim=1) + h = module( + h, + emb, + context=context, + # cam=cam, + image_only_indicator=image_only_indicator, + cond_view=cond_view, + cond_motion=cond_motion, + time_context=time_context, + num_video_frames=num_video_frames, + time_step=time_step, + name='decoder_{}_{}'.format(time, i) + ) + h = h.type(x.dtype) + return self.out(h) diff --git a/sgm/modules/diffusionmodules/wrappers.py b/sgm/modules/diffusionmodules/wrappers.py new file mode 100644 index 0000000..673caab --- /dev/null +++ b/sgm/modules/diffusionmodules/wrappers.py @@ -0,0 +1,73 @@ +import torch +import torch.nn as nn +from packaging import version +from einops import rearrange, repeat + +OPENAIUNETWRAPPER = "sgm.modules.diffusionmodules.wrappers.OpenAIWrapper" + + +class IdentityWrapper(nn.Module): + def __init__(self, diffusion_model, compile_model: bool = False): + super().__init__() + compile = ( + torch.compile + if (version.parse(torch.__version__) >= version.parse("2.0.0")) + and compile_model + else lambda x: x + ) + self.diffusion_model = compile(diffusion_model) + + def forward(self, *args, **kwargs): + return self.diffusion_model(*args, **kwargs) + + +class OpenAIWrapper(IdentityWrapper): + def forward( + self, x: torch.Tensor, t: torch.Tensor, c: dict, **kwargs + ) -> torch.Tensor: + x = torch.cat((x, c.get("concat", torch.Tensor([]).type_as(x))), dim=1) + if "cond_view" in c: + return self.diffusion_model( + x, + timesteps=t, + context=c.get("crossattn", None), + y=c.get("vector", None), + cond_view=c.get("cond_view", None), + cond_motion=c.get("cond_motion", None), + **kwargs, + ) + else: + return self.diffusion_model( + x, + timesteps=t, + context=c.get("crossattn", None), + y=c.get("vector", None), + **kwargs, + ) + +class SevaWrapper(IdentityWrapper): + def forward( + self, x: torch.Tensor, t: torch.Tensor, c: dict, **kwargs + ) -> torch.Tensor: + x = torch.cat((x, c.get("concat", torch.Tensor([]).type_as(x))), dim=2) + + + b = x.shape[0] + f = x.shape[1] + x = rearrange(x, "b f c h w -> (b f) c h w") + dense_y=rearrange(c["plucker"], "b f c h w -> (b f) c h w") + + t = repeat(t, "b -> (b f)", f=f) + y = repeat(c["crossattn"], "b 1 c -> (b f) 1 c", f=f) + + + out = self.diffusion_model( + x, + t=t, + y=y, + dense_y=dense_y, + num_frames=f, + **kwargs, + ) + out = rearrange(out, "(b f) c h w -> b f c h w", f=f) + return out \ No newline at end of file diff --git a/sgm/modules/distributions/__init__.py b/sgm/modules/distributions/__init__.py new file mode 100644 index 0000000..e69de29 diff --git a/sgm/modules/distributions/distributions.py b/sgm/modules/distributions/distributions.py new file mode 100644 index 0000000..016be35 --- /dev/null +++ b/sgm/modules/distributions/distributions.py @@ -0,0 +1,102 @@ +import numpy as np +import torch + + +class AbstractDistribution: + def sample(self): + raise NotImplementedError() + + def mode(self): + raise NotImplementedError() + + +class DiracDistribution(AbstractDistribution): + def __init__(self, value): + self.value = value + + def sample(self): + return self.value + + def mode(self): + return self.value + + +class DiagonalGaussianDistribution(object): + def __init__(self, parameters, deterministic=False): + self.parameters = parameters + self.mean, self.logvar = torch.chunk(parameters, 2, dim=1) + self.logvar = torch.clamp(self.logvar, -30.0, 20.0) + self.deterministic = deterministic + self.std = torch.exp(0.5 * self.logvar) + self.var = torch.exp(self.logvar) + if self.deterministic: + self.var = self.std = torch.zeros_like(self.mean).to( + device=self.parameters.device + ) + + def sample(self): + x = self.mean + self.std * torch.randn(self.mean.shape).to( + device=self.parameters.device + ) + return x + + def kl(self, other=None): + if self.deterministic: + return torch.Tensor([0.0]) + else: + if other is None: + return 0.5 * torch.sum( + torch.pow(self.mean, 2) + self.var - 1.0 - self.logvar, + dim=[1, 2, 3], + ) + else: + return 0.5 * torch.sum( + torch.pow(self.mean - other.mean, 2) / other.var + + self.var / other.var + - 1.0 + - self.logvar + + other.logvar, + dim=[1, 2, 3], + ) + + def nll(self, sample, dims=[1, 2, 3]): + if self.deterministic: + return torch.Tensor([0.0]) + logtwopi = np.log(2.0 * np.pi) + return 0.5 * torch.sum( + logtwopi + self.logvar + torch.pow(sample - self.mean, 2) / self.var, + dim=dims, + ) + + def mode(self): + return self.mean + + +def normal_kl(mean1, logvar1, mean2, logvar2): + """ + source: https://github.com/openai/guided-diffusion/blob/27c20a8fab9cb472df5d6bdd6c8d11c8f430b924/guided_diffusion/losses.py#L12 + Compute the KL divergence between two gaussians. + Shapes are automatically broadcasted, so batches can be compared to + scalars, among other use cases. + """ + tensor = None + for obj in (mean1, logvar1, mean2, logvar2): + if isinstance(obj, torch.Tensor): + tensor = obj + break + assert tensor is not None, "at least one argument must be a Tensor" + + # Force variances to be Tensors. Broadcasting helps convert scalars to + # Tensors, but it does not work for torch.exp(). + logvar1, logvar2 = [ + x if isinstance(x, torch.Tensor) else torch.tensor(x).to(tensor) + for x in (logvar1, logvar2) + ] + + return 0.5 * ( + -1.0 + + logvar2 + - logvar1 + + torch.exp(logvar1 - logvar2) + + ((mean1 - mean2) ** 2) * torch.exp(-logvar2) + ) diff --git a/sgm/modules/ema.py b/sgm/modules/ema.py new file mode 100644 index 0000000..97b5ae2 --- /dev/null +++ b/sgm/modules/ema.py @@ -0,0 +1,86 @@ +import torch +from torch import nn + + +class LitEma(nn.Module): + def __init__(self, model, decay=0.9999, use_num_upates=True): + super().__init__() + if decay < 0.0 or decay > 1.0: + raise ValueError("Decay must be between 0 and 1") + + self.m_name2s_name = {} + self.register_buffer("decay", torch.tensor(decay, dtype=torch.float32)) + self.register_buffer( + "num_updates", + torch.tensor(0, dtype=torch.int) + if use_num_upates + else torch.tensor(-1, dtype=torch.int), + ) + + for name, p in model.named_parameters(): + if p.requires_grad: + # remove as '.'-character is not allowed in buffers + s_name = name.replace(".", "") + self.m_name2s_name.update({name: s_name}) + self.register_buffer(s_name, p.clone().detach().data) + + self.collected_params = [] + + def reset_num_updates(self): + del self.num_updates + self.register_buffer("num_updates", torch.tensor(0, dtype=torch.int)) + + def forward(self, model): + decay = self.decay + + if self.num_updates >= 0: + self.num_updates += 1 + decay = min(self.decay, (1 + self.num_updates) / (10 + self.num_updates)) + + one_minus_decay = 1.0 - decay + + with torch.no_grad(): + m_param = dict(model.named_parameters()) + shadow_params = dict(self.named_buffers()) + + for key in m_param: + if m_param[key].requires_grad: + sname = self.m_name2s_name[key] + shadow_params[sname] = shadow_params[sname].type_as(m_param[key]) + shadow_params[sname].sub_( + one_minus_decay * (shadow_params[sname] - m_param[key]) + ) + else: + assert not key in self.m_name2s_name + + def copy_to(self, model): + m_param = dict(model.named_parameters()) + shadow_params = dict(self.named_buffers()) + for key in m_param: + if m_param[key].requires_grad: + m_param[key].data.copy_(shadow_params[self.m_name2s_name[key]].data) + else: + assert not key in self.m_name2s_name + + def store(self, parameters): + """ + Save the current parameters for restoring later. + Args: + parameters: Iterable of `torch.nn.Parameter`; the parameters to be + temporarily stored. + """ + self.collected_params = [param.clone() for param in parameters] + + def restore(self, parameters): + """ + Restore the parameters stored with the `store` method. + Useful to validate the model with EMA parameters without affecting the + original optimization process. Store the parameters before the + `copy_to` method. After validation (or model saving), use this to + restore the former parameters. + Args: + parameters: Iterable of `torch.nn.Parameter`; the parameters to be + updated with the stored parameters. + """ + for c_param, param in zip(self.collected_params, parameters): + param.data.copy_(c_param.data) diff --git a/sgm/modules/encoders/__init__.py b/sgm/modules/encoders/__init__.py new file mode 100644 index 0000000..e69de29 diff --git a/sgm/modules/encoders/modules.py b/sgm/modules/encoders/modules.py new file mode 100644 index 0000000..24ee41a --- /dev/null +++ b/sgm/modules/encoders/modules.py @@ -0,0 +1,1140 @@ +import math +from contextlib import nullcontext +from functools import partial +from typing import Dict, List, Optional, Tuple, Union + +import kornia +import numpy as np +import open_clip +import torch +import torch.nn as nn +from einops import rearrange, repeat +from omegaconf import ListConfig +from torch.utils.checkpoint import checkpoint +from transformers import (ByT5Tokenizer, CLIPTextModel, CLIPTokenizer, + T5EncoderModel, T5Tokenizer) + +from ...modules.autoencoding.regularizers import DiagonalGaussianRegularizer +from ...modules.diffusionmodules.model import Encoder +from ...modules.diffusionmodules.openaimodel import Timestep +from ...modules.diffusionmodules.util import (extract_into_tensor, + make_beta_schedule) +from ...modules.distributions.distributions import DiagonalGaussianDistribution +from ...util import (append_dims, autocast, count_params, default, + disabled_train, expand_dims_like, instantiate_from_config) + + +class AbstractEmbModel(nn.Module): + def __init__(self): + super().__init__() + self._is_trainable = None + self._ucg_rate = None + self._input_key = None + + @property + def is_trainable(self) -> bool: + return self._is_trainable + + @property + def ucg_rate(self) -> Union[float, torch.Tensor]: + return self._ucg_rate + + @property + def input_key(self) -> str: + return self._input_key + + @is_trainable.setter + def is_trainable(self, value: bool): + self._is_trainable = value + + @ucg_rate.setter + def ucg_rate(self, value: Union[float, torch.Tensor]): + self._ucg_rate = value + + @input_key.setter + def input_key(self, value: str): + self._input_key = value + + @is_trainable.deleter + def is_trainable(self): + del self._is_trainable + + @ucg_rate.deleter + def ucg_rate(self): + del self._ucg_rate + + @input_key.deleter + def input_key(self): + del self._input_key + + +class GeneralConditioner(nn.Module): + OUTPUT_DIM2KEYS = {2: "vector", 3: "crossattn", 4: "concat", 5: "concat"} + KEY2CATDIM = {"vector": 1, "crossattn": 2, "concat": 1, "cond_view": 1, "cond_motion": 1} + + def __init__(self, emb_models: Union[List, ListConfig]): + super().__init__() + embedders = [] + for n, embconfig in enumerate(emb_models): + embedder = instantiate_from_config(embconfig) + assert isinstance( + embedder, AbstractEmbModel + ), f"embedder model {embedder.__class__.__name__} has to inherit from AbstractEmbModel" + embedder.is_trainable = embconfig.get("is_trainable", False) + embedder.ucg_rate = embconfig.get("ucg_rate", 0.0) + if not embedder.is_trainable: + embedder.train = disabled_train + for param in embedder.parameters(): + param.requires_grad = False + embedder.eval() + print( + f"Initialized embedder #{n}: {embedder.__class__.__name__} " + f"with {count_params(embedder, False)} params. Trainable: {embedder.is_trainable}" + ) + + if "input_key" in embconfig: + embedder.input_key = embconfig["input_key"] + elif "input_keys" in embconfig: + embedder.input_keys = embconfig["input_keys"] + else: + raise KeyError( + f"need either 'input_key' or 'input_keys' for embedder {embedder.__class__.__name__}" + ) + + embedder.legacy_ucg_val = embconfig.get("legacy_ucg_value", None) + if embedder.legacy_ucg_val is not None: + embedder.ucg_prng = np.random.RandomState() + + embedders.append(embedder) + self.embedders = nn.ModuleList(embedders) + + def possibly_get_ucg_val(self, embedder: AbstractEmbModel, batch: Dict) -> Dict: + assert embedder.legacy_ucg_val is not None + p = embedder.ucg_rate + val = embedder.legacy_ucg_val + for i in range(len(batch[embedder.input_key])): + if embedder.ucg_prng.choice(2, p=[1 - p, p]): + batch[embedder.input_key][i] = val + return batch + + def forward( + self, batch: Dict, force_zero_embeddings: Optional[List] = None + ) -> Dict: + output = dict() + if force_zero_embeddings is None: + force_zero_embeddings = [] + for embedder in self.embedders: + embedding_context = nullcontext if embedder.is_trainable else torch.no_grad + with embedding_context(): + if hasattr(embedder, "input_key") and (embedder.input_key is not None): + if embedder.legacy_ucg_val is not None: + batch = self.possibly_get_ucg_val(embedder, batch) + emb_out = embedder(batch[embedder.input_key]) + elif hasattr(embedder, "input_keys"): + emb_out = embedder(*[batch[k] for k in embedder.input_keys]) + assert isinstance( + emb_out, (torch.Tensor, list, tuple) + ), f"encoder outputs must be tensors or a sequence, but got {type(emb_out)}" + if not isinstance(emb_out, (list, tuple)): + emb_out = [emb_out] + for emb in emb_out: + if embedder.input_key in ["cond_view", "cond_motion", "plucker", "mask", "replace"]: + out_key = embedder.input_key + else: + out_key = self.OUTPUT_DIM2KEYS[emb.dim()] + + if embedder.ucg_rate > 0.0 and embedder.legacy_ucg_val is None: + emb = ( + expand_dims_like( + torch.bernoulli( + (1.0 - embedder.ucg_rate) + * torch.ones(emb.shape[0], device=emb.device) + ), + emb, + ) + * emb + ) + if ( + hasattr(embedder, "input_key") + and embedder.input_key in force_zero_embeddings + ): + emb = torch.zeros_like(emb) + if out_key in output: + output[out_key] = torch.cat( + (output[out_key], emb), self.KEY2CATDIM[out_key] + ) + else: + output[out_key] = emb + return output + + def get_unconditional_conditioning( + self, + batch_c: Dict, + batch_uc: Optional[Dict] = None, + force_uc_zero_embeddings: Optional[List[str]] = None, + force_cond_zero_embeddings: Optional[List[str]] = None, + ): + if force_uc_zero_embeddings is None: + force_uc_zero_embeddings = [] + ucg_rates = list() + for embedder in self.embedders: + ucg_rates.append(embedder.ucg_rate) + embedder.ucg_rate = 0.0 + c = self(batch_c, force_cond_zero_embeddings) + uc = self(batch_c if batch_uc is None else batch_uc, force_uc_zero_embeddings) + + for embedder, rate in zip(self.embedders, ucg_rates): + embedder.ucg_rate = rate + return c, uc + + +class InceptionV3(nn.Module): + """Wrapper around the https://github.com/mseitzer/pytorch-fid inception + port with an additional squeeze at the end""" + + def __init__(self, normalize_input=False, **kwargs): + super().__init__() + from pytorch_fid import inception + + kwargs["resize_input"] = True + self.model = inception.InceptionV3(normalize_input=normalize_input, **kwargs) + + def forward(self, inp): + outp = self.model(inp) + + if len(outp) == 1: + return outp[0].squeeze() + + return outp + + +class IdentityEncoder(AbstractEmbModel): + def encode(self, x): + return x + + def forward(self, x): + return x + + +class ClassEmbedder(AbstractEmbModel): + def __init__(self, embed_dim, n_classes=1000, add_sequence_dim=False): + super().__init__() + self.embedding = nn.Embedding(n_classes, embed_dim) + self.n_classes = n_classes + self.add_sequence_dim = add_sequence_dim + + def forward(self, c): + c = self.embedding(c) + if self.add_sequence_dim: + c = c[:, None, :] + return c + + def get_unconditional_conditioning(self, bs, device="cuda"): + uc_class = ( + self.n_classes - 1 + ) # 1000 classes --> 0 ... 999, one extra class for ucg (class 1000) + uc = torch.ones((bs,), device=device) * uc_class + uc = {self.key: uc.long()} + return uc + + +class ClassEmbedderForMultiCond(ClassEmbedder): + def forward(self, batch, key=None, disable_dropout=False): + out = batch + key = default(key, self.key) + islist = isinstance(batch[key], list) + if islist: + batch[key] = batch[key][0] + c_out = super().forward(batch, key, disable_dropout) + out[key] = [c_out] if islist else c_out + return out + + +class FrozenT5Embedder(AbstractEmbModel): + """Uses the T5 transformer encoder for text""" + + def __init__( + self, version="google/t5-v1_1-xxl", device="cuda", max_length=77, freeze=True + ): # others are google/t5-v1_1-xl and google/t5-v1_1-xxl + super().__init__() + self.tokenizer = T5Tokenizer.from_pretrained(version) + self.transformer = T5EncoderModel.from_pretrained(version) + self.device = device + self.max_length = max_length + if freeze: + self.freeze() + + def freeze(self): + self.transformer = self.transformer.eval() + + for param in self.parameters(): + param.requires_grad = False + + def forward(self, text): + batch_encoding = self.tokenizer( + text, + truncation=True, + max_length=self.max_length, + return_length=True, + return_overflowing_tokens=False, + padding="max_length", + return_tensors="pt", + ) + tokens = batch_encoding["input_ids"].to(self.device) + with torch.autocast("cuda", enabled=False): + outputs = self.transformer(input_ids=tokens) + z = outputs.last_hidden_state + return z + + def encode(self, text): + return self(text) + + +class FrozenByT5Embedder(AbstractEmbModel): + """ + Uses the ByT5 transformer encoder for text. Is character-aware. + """ + + def __init__( + self, version="google/byt5-base", device="cuda", max_length=77, freeze=True + ): # others are google/t5-v1_1-xl and google/t5-v1_1-xxl + super().__init__() + self.tokenizer = ByT5Tokenizer.from_pretrained(version) + self.transformer = T5EncoderModel.from_pretrained(version) + self.device = device + self.max_length = max_length + if freeze: + self.freeze() + + def freeze(self): + self.transformer = self.transformer.eval() + + for param in self.parameters(): + param.requires_grad = False + + def forward(self, text): + batch_encoding = self.tokenizer( + text, + truncation=True, + max_length=self.max_length, + return_length=True, + return_overflowing_tokens=False, + padding="max_length", + return_tensors="pt", + ) + tokens = batch_encoding["input_ids"].to(self.device) + with torch.autocast("cuda", enabled=False): + outputs = self.transformer(input_ids=tokens) + z = outputs.last_hidden_state + return z + + def encode(self, text): + return self(text) + + +class FrozenCLIPEmbedder(AbstractEmbModel): + """Uses the CLIP transformer encoder for text (from huggingface)""" + + LAYERS = ["last", "pooled", "hidden"] + + def __init__( + self, + version="openai/clip-vit-large-patch14", + device="cuda", + max_length=77, + freeze=True, + layer="last", + layer_idx=None, + always_return_pooled=False, + ): # clip-vit-base-patch32 + super().__init__() + assert layer in self.LAYERS + self.tokenizer = CLIPTokenizer.from_pretrained(version) + self.transformer = CLIPTextModel.from_pretrained(version) + self.device = device + self.max_length = max_length + if freeze: + self.freeze() + self.layer = layer + self.layer_idx = layer_idx + self.return_pooled = always_return_pooled + if layer == "hidden": + assert layer_idx is not None + assert 0 <= abs(layer_idx) <= 12 + + def freeze(self): + self.transformer = self.transformer.eval() + + for param in self.parameters(): + param.requires_grad = False + + @autocast + def forward(self, text): + batch_encoding = self.tokenizer( + text, + truncation=True, + max_length=self.max_length, + return_length=True, + return_overflowing_tokens=False, + padding="max_length", + return_tensors="pt", + ) + tokens = batch_encoding["input_ids"].to(self.device) + outputs = self.transformer( + input_ids=tokens, output_hidden_states=self.layer == "hidden" + ) + if self.layer == "last": + z = outputs.last_hidden_state + elif self.layer == "pooled": + z = outputs.pooler_output[:, None, :] + else: + z = outputs.hidden_states[self.layer_idx] + if self.return_pooled: + return z, outputs.pooler_output + return z + + def encode(self, text): + return self(text) + + +class FrozenOpenCLIPEmbedder2(AbstractEmbModel): + """ + Uses the OpenCLIP transformer encoder for text + """ + + LAYERS = ["pooled", "last", "penultimate"] + + def __init__( + self, + arch="ViT-H-14", + version="laion2b_s32b_b79k", + device="cuda", + max_length=77, + freeze=True, + layer="last", + always_return_pooled=False, + legacy=True, + ): + super().__init__() + assert layer in self.LAYERS + model, _, _ = open_clip.create_model_and_transforms( + arch, + device=torch.device("cpu"), + pretrained=version, + ) + del model.visual + self.model = model + + self.device = device + self.max_length = max_length + self.return_pooled = always_return_pooled + if freeze: + self.freeze() + self.layer = layer + if self.layer == "last": + self.layer_idx = 0 + elif self.layer == "penultimate": + self.layer_idx = 1 + else: + raise NotImplementedError() + self.legacy = legacy + + def freeze(self): + self.model = self.model.eval() + for param in self.parameters(): + param.requires_grad = False + + @autocast + def forward(self, text): + tokens = open_clip.tokenize(text) + z = self.encode_with_transformer(tokens.to(self.device)) + if not self.return_pooled and self.legacy: + return z + if self.return_pooled: + assert not self.legacy + return z[self.layer], z["pooled"] + return z[self.layer] + + def encode_with_transformer(self, text): + x = self.model.token_embedding(text) # [batch_size, n_ctx, d_model] + x = x + self.model.positional_embedding + x = x.permute(1, 0, 2) # NLD -> LND + x = self.text_transformer_forward(x, attn_mask=self.model.attn_mask) + if self.legacy: + x = x[self.layer] + x = self.model.ln_final(x) + return x + else: + # x is a dict and will stay a dict + o = x["last"] + o = self.model.ln_final(o) + pooled = self.pool(o, text) + x["pooled"] = pooled + return x + + def pool(self, x, text): + # take features from the eot embedding (eot_token is the highest number in each sequence) + x = ( + x[torch.arange(x.shape[0]), text.argmax(dim=-1)] + @ self.model.text_projection + ) + return x + + def text_transformer_forward(self, x: torch.Tensor, attn_mask=None): + outputs = {} + for i, r in enumerate(self.model.transformer.resblocks): + if i == len(self.model.transformer.resblocks) - 1: + outputs["penultimate"] = x.permute(1, 0, 2) # LND -> NLD + if ( + self.model.transformer.grad_checkpointing + and not torch.jit.is_scripting() + ): + x = checkpoint(r, x, attn_mask) + else: + x = r(x, attn_mask=attn_mask) + outputs["last"] = x.permute(1, 0, 2) # LND -> NLD + return outputs + + def encode(self, text): + return self(text) + + +class FrozenOpenCLIPEmbedder(AbstractEmbModel): + LAYERS = [ + # "pooled", + "last", + "penultimate", + ] + + def __init__( + self, + arch="ViT-H-14", + version="laion2b_s32b_b79k", + device="cuda", + max_length=77, + freeze=True, + layer="last", + ): + super().__init__() + assert layer in self.LAYERS + model, _, _ = open_clip.create_model_and_transforms( + arch, device=torch.device("cpu"), pretrained=version + ) + del model.visual + self.model = model + + self.device = device + self.max_length = max_length + if freeze: + self.freeze() + self.layer = layer + if self.layer == "last": + self.layer_idx = 0 + elif self.layer == "penultimate": + self.layer_idx = 1 + else: + raise NotImplementedError() + + def freeze(self): + self.model = self.model.eval() + for param in self.parameters(): + param.requires_grad = False + + def forward(self, text): + tokens = open_clip.tokenize(text) + z = self.encode_with_transformer(tokens.to(self.device)) + return z + + def encode_with_transformer(self, text): + x = self.model.token_embedding(text) # [batch_size, n_ctx, d_model] + x = x + self.model.positional_embedding + x = x.permute(1, 0, 2) # NLD -> LND + x = self.text_transformer_forward(x, attn_mask=self.model.attn_mask) + x = x.permute(1, 0, 2) # LND -> NLD + x = self.model.ln_final(x) + return x + + def text_transformer_forward(self, x: torch.Tensor, attn_mask=None): + for i, r in enumerate(self.model.transformer.resblocks): + if i == len(self.model.transformer.resblocks) - self.layer_idx: + break + if ( + self.model.transformer.grad_checkpointing + and not torch.jit.is_scripting() + ): + x = checkpoint(r, x, attn_mask) + else: + x = r(x, attn_mask=attn_mask) + return x + + def encode(self, text): + return self(text) + + +class FrozenOpenCLIPImageEmbedder(AbstractEmbModel): + """ + Uses the OpenCLIP vision transformer encoder for images + """ + + def __init__( + self, + arch="ViT-H-14", + version="laion2b_s32b_b79k", + device="cuda", + max_length=77, + freeze=True, + antialias=True, + ucg_rate=0.0, + unsqueeze_dim=False, + repeat_to_max_len=False, + num_image_crops=0, + output_tokens=False, + init_device=None, + ): + super().__init__() + model, _, _ = open_clip.create_model_and_transforms( + arch, + device=torch.device(default(init_device, "cpu")), + pretrained=version, + ) + del model.transformer + self.model = model + self.max_crops = num_image_crops + self.pad_to_max_len = self.max_crops > 0 + self.repeat_to_max_len = repeat_to_max_len and (not self.pad_to_max_len) + self.device = device + self.max_length = max_length + if freeze: + self.freeze() + + self.antialias = antialias + + self.register_buffer( + "mean", torch.Tensor([0.48145466, 0.4578275, 0.40821073]), persistent=False + ) + self.register_buffer( + "std", torch.Tensor([0.26862954, 0.26130258, 0.27577711]), persistent=False + ) + self.ucg_rate = ucg_rate + self.unsqueeze_dim = unsqueeze_dim + self.stored_batch = None + self.model.visual.output_tokens = output_tokens + self.output_tokens = output_tokens + + def preprocess(self, x): + # normalize to [0,1] + x = kornia.geometry.resize( + x, + (224, 224), + interpolation="bicubic", + align_corners=True, + antialias=self.antialias, + ) + x = (x + 1.0) / 2.0 + # renormalize according to clip + x = kornia.enhance.normalize(x, self.mean, self.std) + return x + + def freeze(self): + self.model = self.model.eval() + for param in self.parameters(): + param.requires_grad = False + + @autocast + def forward(self, image, no_dropout=False): + z = self.encode_with_vision_transformer(image) + tokens = None + if self.output_tokens: + z, tokens = z[0], z[1] + z = z.to(image.dtype) + if self.ucg_rate > 0.0 and not no_dropout and not (self.max_crops > 0): + z = ( + torch.bernoulli( + (1.0 - self.ucg_rate) * torch.ones(z.shape[0], device=z.device) + )[:, None] + * z + ) + if tokens is not None: + tokens = ( + expand_dims_like( + torch.bernoulli( + (1.0 - self.ucg_rate) + * torch.ones(tokens.shape[0], device=tokens.device) + ), + tokens, + ) + * tokens + ) + if self.unsqueeze_dim: + z = z[:, None, :] + if self.output_tokens: + assert not self.repeat_to_max_len + assert not self.pad_to_max_len + return tokens, z + if self.repeat_to_max_len: + if z.dim() == 2: + z_ = z[:, None, :] + else: + z_ = z + return repeat(z_, "b 1 d -> b n d", n=self.max_length), z + elif self.pad_to_max_len: + assert z.dim() == 3 + z_pad = torch.cat( + ( + z, + torch.zeros( + z.shape[0], + self.max_length - z.shape[1], + z.shape[2], + device=z.device, + ), + ), + 1, + ) + return z_pad, z_pad[:, 0, ...] + return z + + def encode_with_vision_transformer(self, img): + # if self.max_crops > 0: + # img = self.preprocess_by_cropping(img) + if img.dim() == 5: + assert self.max_crops == img.shape[1] + img = rearrange(img, "b n c h w -> (b n) c h w") + img = self.preprocess(img) + if not self.output_tokens: + assert not self.model.visual.output_tokens + x = self.model.visual(img) + tokens = None + else: + assert self.model.visual.output_tokens + x, tokens = self.model.visual(img) + if self.max_crops > 0: + x = rearrange(x, "(b n) d -> b n d", n=self.max_crops) + # drop out between 0 and all along the sequence axis + x = ( + torch.bernoulli( + (1.0 - self.ucg_rate) + * torch.ones(x.shape[0], x.shape[1], 1, device=x.device) + ) + * x + ) + if tokens is not None: + tokens = rearrange(tokens, "(b n) t d -> b t (n d)", n=self.max_crops) + print( + f"You are running very experimental token-concat in {self.__class__.__name__}. " + f"Check what you are doing, and then remove this message." + ) + if self.output_tokens: + return x, tokens + return x + + def encode(self, text): + return self(text) + + +class FrozenCLIPT5Encoder(AbstractEmbModel): + def __init__( + self, + clip_version="openai/clip-vit-large-patch14", + t5_version="google/t5-v1_1-xl", + device="cuda", + clip_max_length=77, + t5_max_length=77, + ): + super().__init__() + self.clip_encoder = FrozenCLIPEmbedder( + clip_version, device, max_length=clip_max_length + ) + self.t5_encoder = FrozenT5Embedder(t5_version, device, max_length=t5_max_length) + print( + f"{self.clip_encoder.__class__.__name__} has {count_params(self.clip_encoder) * 1.e-6:.2f} M parameters, " + f"{self.t5_encoder.__class__.__name__} comes with {count_params(self.t5_encoder) * 1.e-6:.2f} M params." + ) + + def encode(self, text): + return self(text) + + def forward(self, text): + clip_z = self.clip_encoder.encode(text) + t5_z = self.t5_encoder.encode(text) + return [clip_z, t5_z] + + +class SpatialRescaler(nn.Module): + def __init__( + self, + n_stages=1, + method="bilinear", + multiplier=0.5, + in_channels=3, + out_channels=None, + bias=False, + wrap_video=False, + kernel_size=1, + remap_output=False, + ): + super().__init__() + self.n_stages = n_stages + assert self.n_stages >= 0 + assert method in [ + "nearest", + "linear", + "bilinear", + "trilinear", + "bicubic", + "area", + ] + self.multiplier = multiplier + self.interpolator = partial(torch.nn.functional.interpolate, mode=method) + self.remap_output = out_channels is not None or remap_output + if self.remap_output: + print( + f"Spatial Rescaler mapping from {in_channels} to {out_channels} channels after resizing." + ) + self.channel_mapper = nn.Conv2d( + in_channels, + out_channels, + kernel_size=kernel_size, + bias=bias, + padding=kernel_size // 2, + ) + self.wrap_video = wrap_video + + def forward(self, x): + if self.wrap_video and x.ndim == 5: + B, C, T, H, W = x.shape + x = rearrange(x, "b c t h w -> b t c h w") + x = rearrange(x, "b t c h w -> (b t) c h w") + + for stage in range(self.n_stages): + x = self.interpolator(x, scale_factor=self.multiplier) + + if self.wrap_video: + x = rearrange(x, "(b t) c h w -> b t c h w", b=B, t=T, c=C) + x = rearrange(x, "b t c h w -> b c t h w") + if self.remap_output: + x = self.channel_mapper(x) + return x + + def encode(self, x): + return self(x) + + +class LowScaleEncoder(nn.Module): + def __init__( + self, + model_config, + linear_start, + linear_end, + timesteps=1000, + max_noise_level=250, + output_size=64, + scale_factor=1.0, + ): + super().__init__() + self.max_noise_level = max_noise_level + self.model = instantiate_from_config(model_config) + self.augmentation_schedule = self.register_schedule( + timesteps=timesteps, linear_start=linear_start, linear_end=linear_end + ) + self.out_size = output_size + self.scale_factor = scale_factor + + def register_schedule( + self, + beta_schedule="linear", + timesteps=1000, + linear_start=1e-4, + linear_end=2e-2, + cosine_s=8e-3, + ): + betas = make_beta_schedule( + beta_schedule, + timesteps, + linear_start=linear_start, + linear_end=linear_end, + cosine_s=cosine_s, + ) + alphas = 1.0 - betas + alphas_cumprod = np.cumprod(alphas, axis=0) + alphas_cumprod_prev = np.append(1.0, alphas_cumprod[:-1]) + + (timesteps,) = betas.shape + self.num_timesteps = int(timesteps) + self.linear_start = linear_start + self.linear_end = linear_end + assert ( + alphas_cumprod.shape[0] == self.num_timesteps + ), "alphas have to be defined for each timestep" + + to_torch = partial(torch.tensor, dtype=torch.float32) + + self.register_buffer("betas", to_torch(betas)) + self.register_buffer("alphas_cumprod", to_torch(alphas_cumprod)) + self.register_buffer("alphas_cumprod_prev", to_torch(alphas_cumprod_prev)) + + # calculations for diffusion q(x_t | x_{t-1}) and others + self.register_buffer("sqrt_alphas_cumprod", to_torch(np.sqrt(alphas_cumprod))) + self.register_buffer( + "sqrt_one_minus_alphas_cumprod", to_torch(np.sqrt(1.0 - alphas_cumprod)) + ) + self.register_buffer( + "log_one_minus_alphas_cumprod", to_torch(np.log(1.0 - alphas_cumprod)) + ) + self.register_buffer( + "sqrt_recip_alphas_cumprod", to_torch(np.sqrt(1.0 / alphas_cumprod)) + ) + self.register_buffer( + "sqrt_recipm1_alphas_cumprod", to_torch(np.sqrt(1.0 / alphas_cumprod - 1)) + ) + + def q_sample(self, x_start, t, noise=None): + noise = default(noise, lambda: torch.randn_like(x_start)) + return ( + extract_into_tensor(self.sqrt_alphas_cumprod, t, x_start.shape) * x_start + + extract_into_tensor(self.sqrt_one_minus_alphas_cumprod, t, x_start.shape) + * noise + ) + + def forward(self, x): + z = self.model.encode(x) + if isinstance(z, DiagonalGaussianDistribution): + z = z.sample() + z = z * self.scale_factor + noise_level = torch.randint( + 0, self.max_noise_level, (x.shape[0],), device=x.device + ).long() + z = self.q_sample(z, noise_level) + if self.out_size is not None: + z = torch.nn.functional.interpolate(z, size=self.out_size, mode="nearest") + return z, noise_level + + def decode(self, z): + z = z / self.scale_factor + return self.model.decode(z) + + +class ConcatTimestepEmbedderND(AbstractEmbModel): + """embeds each dimension independently and concatenates them""" + + def __init__(self, outdim): + super().__init__() + self.timestep = Timestep(outdim) + self.outdim = outdim + + def forward(self, x): + if x.ndim == 1: + x = x[:, None] + assert len(x.shape) == 2 + b, dims = x.shape[0], x.shape[1] + x = rearrange(x, "b d -> (b d)") + emb = self.timestep(x) + emb = rearrange(emb, "(b d) d2 -> b (d d2)", b=b, d=dims, d2=self.outdim) + return emb + + +class GaussianEncoder(Encoder, AbstractEmbModel): + def __init__( + self, weight: float = 1.0, flatten_output: bool = True, *args, **kwargs + ): + super().__init__(*args, **kwargs) + self.posterior = DiagonalGaussianRegularizer() + self.weight = weight + self.flatten_output = flatten_output + + def forward(self, x) -> Tuple[Dict, torch.Tensor]: + z = super().forward(x) + z, log = self.posterior(z) + log["loss"] = log["kl_loss"] + log["weight"] = self.weight + if self.flatten_output: + z = rearrange(z, "b c h w -> b (h w ) c") + return log, z + + +class VideoPredictionEmbedderWithEncoder(AbstractEmbModel): + def __init__( + self, + n_cond_frames: int, + n_copies: int, + encoder_config: dict, + sigma_sampler_config: Optional[dict] = None, + sigma_cond_config: Optional[dict] = None, + is_ae: bool = False, + scale_factor: float = 1.0, + disable_encoder_autocast: bool = False, + en_and_decode_n_samples_a_time: Optional[int] = None, + ): + super().__init__() + + self.n_cond_frames = n_cond_frames + self.n_copies = n_copies + self.encoder = instantiate_from_config(encoder_config) + self.sigma_sampler = ( + instantiate_from_config(sigma_sampler_config) + if sigma_sampler_config is not None + else None + ) + self.sigma_cond = ( + instantiate_from_config(sigma_cond_config) + if sigma_cond_config is not None + else None + ) + self.is_ae = is_ae + self.scale_factor = scale_factor + self.disable_encoder_autocast = disable_encoder_autocast + self.en_and_decode_n_samples_a_time = en_and_decode_n_samples_a_time + + def forward( + self, vid: torch.Tensor + ) -> Union[ + torch.Tensor, + Tuple[torch.Tensor, torch.Tensor], + Tuple[torch.Tensor, dict], + Tuple[Tuple[torch.Tensor, torch.Tensor], dict], + ]: + if self.sigma_sampler is not None: + b = vid.shape[0] // self.n_cond_frames + sigmas = self.sigma_sampler(b).to(vid.device) + if self.sigma_cond is not None: + sigma_cond = self.sigma_cond(sigmas) + if self.n_cond_frames == 1: + sigma_cond = repeat(sigma_cond, "b d -> (b t) d", t=self.n_copies) + else: + sigma_cond = repeat(sigma_cond, "b d -> (b t) d", t=self.n_cond_frames) # For SV4D + sigmas = repeat(sigmas, "b -> (b t)", t=self.n_cond_frames) + noise = torch.randn_like(vid) + vid = vid + noise * append_dims(sigmas, vid.ndim) + + with torch.autocast("cuda", enabled=not self.disable_encoder_autocast): + n_samples = ( + self.en_and_decode_n_samples_a_time + if self.en_and_decode_n_samples_a_time is not None + else vid.shape[0] + ) + n_rounds = math.ceil(vid.shape[0] / n_samples) + all_out = [] + for n in range(n_rounds): + if self.is_ae: + out = self.encoder.encode(vid[n * n_samples : (n + 1) * n_samples]) + else: + out = self.encoder(vid[n * n_samples : (n + 1) * n_samples]) + all_out.append(out) + + vid = torch.cat(all_out, dim=0) + vid *= self.scale_factor + + if self.n_cond_frames == 1: + vid = rearrange(vid, "(b t) c h w -> b () (t c) h w", t=self.n_cond_frames) + vid = repeat(vid, "b 1 c h w -> (b t) c h w", t=self.n_copies) + + return_val = (vid, sigma_cond) if self.sigma_cond is not None else vid + + return return_val + + +class FrozenOpenCLIPImagePredictionEmbedder(AbstractEmbModel): + def __init__( + self, + open_clip_embedding_config: Dict, + n_cond_frames: int, + n_copies: int, + ): + super().__init__() + + self.n_cond_frames = n_cond_frames + self.n_copies = n_copies + self.open_clip = instantiate_from_config(open_clip_embedding_config) + + def forward(self, vid): + vid = self.open_clip(vid) + vid = rearrange(vid, "(b t) d -> b t d", t=self.n_cond_frames) + vid = repeat(vid, "b t d -> (b s) t d", s=self.n_copies) + + return vid + + +class SevaFrozenOpenCLIPImageEmbedder(AbstractEmbModel): + """ + Uses the OpenCLIP vision transformer encoder for images + """ + + def __init__( + self, + arch="ViT-H-14", + version="laion2b_s32b_b79k", + device="cuda", + max_length=77, + freeze=True, + antialias=True, + ucg_rate=0.0, + unsqueeze_dim=False, + init_device=None, + ): + super().__init__() + model, _, _ = open_clip.create_model_and_transforms( + arch, + device=torch.device(default(init_device, "cpu")), + pretrained=version, + ) + del model.transformer + self.model = model + self.device = device + self.max_length = max_length + if freeze: + self.freeze() + + self.antialias = antialias + + self.register_buffer( + "mean", torch.Tensor([0.48145466, 0.4578275, 0.40821073]), persistent=False + ) + self.register_buffer( + "std", torch.Tensor([0.26862954, 0.26130258, 0.27577711]), persistent=False + ) + self.ucg_rate = ucg_rate + self.stored_batch = None + + def preprocess(self, x): + # normalize to [0,1] + x = kornia.geometry.resize( + x, + (224, 224), + interpolation="bicubic", + align_corners=True, + antialias=self.antialias, + ) + x = (x + 1.0) / 2.0 + # renormalize according to clip + x = kornia.enhance.normalize(x, self.mean, self.std) + return x + + def freeze(self): + self.model = self.model.eval() + for param in self.parameters(): + param.requires_grad = False + + @autocast + def forward(self, image, mask, no_dropout=False): + batch_size = image.shape[0] + z = [self.encode_with_vision_transformer(image[b][mask[b]]).mean(0, keepdim=True) + for b in range(batch_size)] + z = torch.cat(z, dim=0) + z = z.to(image.dtype) + if self.ucg_rate > 0.0 and not no_dropout: + z = ( + torch.bernoulli( + (1.0 - self.ucg_rate) * torch.ones(z.shape[0], device=z.device) + )[:, None] + * z + ) + z = z[:, None] + return z + + def encode_with_vision_transformer(self, img): + img = self.preprocess(img) + x = self.model.visual(img) + return x + + def encode(self, text): + return self(text) \ No newline at end of file diff --git a/sgm/modules/spacetime_attention.py b/sgm/modules/spacetime_attention.py new file mode 100644 index 0000000..1ca44bb --- /dev/null +++ b/sgm/modules/spacetime_attention.py @@ -0,0 +1,596 @@ +from functools import partial + +import torch + +from ..modules.attention import * +from ..modules.diffusionmodules.util import ( + AlphaBlender, + get_alpha, + linear, + mixed_checkpoint, + timestep_embedding, +) + + +class TimeMixSequential(nn.Sequential): + def forward(self, x, context=None, timesteps=None): + for layer in self: + x = layer(x, context, timesteps) + + return x + + +class BasicTransformerTimeMixBlock(nn.Module): + ATTENTION_MODES = { + "softmax": CrossAttention, + "softmax-xformers": MemoryEfficientCrossAttention, + } + + def __init__( + self, + dim, + n_heads, + d_head, + dropout=0.0, + context_dim=None, + gated_ff=True, + checkpoint=True, + timesteps=None, + ff_in=False, + inner_dim=None, + attn_mode="softmax", + disable_self_attn=False, + disable_temporal_crossattention=False, + switch_temporal_ca_to_sa=False, + ): + super().__init__() + + attn_cls = self.ATTENTION_MODES[attn_mode] + + self.ff_in = ff_in or inner_dim is not None + if inner_dim is None: + inner_dim = dim + + assert int(n_heads * d_head) == inner_dim + + self.is_res = inner_dim == dim + + if self.ff_in: + self.norm_in = nn.LayerNorm(dim) + self.ff_in = FeedForward( + dim, dim_out=inner_dim, dropout=dropout, glu=gated_ff + ) + + self.timesteps = timesteps + self.disable_self_attn = disable_self_attn + if self.disable_self_attn: + self.attn1 = attn_cls( + query_dim=inner_dim, + heads=n_heads, + dim_head=d_head, + context_dim=context_dim, + dropout=dropout, + ) # is a cross-attention + else: + self.attn1 = attn_cls( + query_dim=inner_dim, heads=n_heads, dim_head=d_head, dropout=dropout + ) # is a self-attention + + self.ff = FeedForward(inner_dim, dim_out=dim, dropout=dropout, glu=gated_ff) + + if disable_temporal_crossattention: + if switch_temporal_ca_to_sa: + raise ValueError + else: + self.attn2 = None + else: + self.norm2 = nn.LayerNorm(inner_dim) + if switch_temporal_ca_to_sa: + self.attn2 = attn_cls( + query_dim=inner_dim, heads=n_heads, dim_head=d_head, dropout=dropout + ) # is a self-attention + else: + self.attn2 = attn_cls( + query_dim=inner_dim, + context_dim=context_dim, + heads=n_heads, + dim_head=d_head, + dropout=dropout, + ) # is self-attn if context is none + + self.norm1 = nn.LayerNorm(inner_dim) + self.norm3 = nn.LayerNorm(inner_dim) + self.switch_temporal_ca_to_sa = switch_temporal_ca_to_sa + + self.checkpoint = checkpoint + if self.checkpoint: + logpy.info(f"{self.__class__.__name__} is using checkpointing") + + def forward( + self, x: torch.Tensor, context: torch.Tensor = None, timesteps: int = None + ) -> torch.Tensor: + if self.checkpoint: + return checkpoint(self._forward, x, context, timesteps) + else: + return self._forward(x, context, timesteps=timesteps) + + def _forward(self, x, context=None, timesteps=None): + assert self.timesteps or timesteps + assert not (self.timesteps and timesteps) or self.timesteps == timesteps + timesteps = self.timesteps or timesteps + B, S, C = x.shape + x = rearrange(x, "(b t) s c -> (b s) t c", t=timesteps) + + if self.ff_in: + x_skip = x + x = self.ff_in(self.norm_in(x)) + if self.is_res: + x += x_skip + + if self.disable_self_attn: + x = self.attn1(self.norm1(x), context=context) + x + else: + x = self.attn1(self.norm1(x)) + x + + if self.attn2 is not None: + if self.switch_temporal_ca_to_sa: + x = self.attn2(self.norm2(x)) + x + else: + x = self.attn2(self.norm2(x), context=context) + x + x_skip = x + x = self.ff(self.norm3(x)) + if self.is_res: + x += x_skip + + x = rearrange( + x, "(b s) t c -> (b t) s c", s=S, b=B // timesteps, c=C, t=timesteps + ) + return x + + def get_last_layer(self): + return self.ff.net[-1].weight + + +class PostHocSpatialTransformerWithTimeMixing(SpatialTransformer): + def __init__( + self, + in_channels, + n_heads, + d_head, + depth=1, + dropout=0.0, + use_linear=False, + context_dim=None, + use_spatial_context=False, + timesteps=None, + merge_strategy: str = "fixed", + merge_factor: float = 0.5, + apply_sigmoid_to_merge: bool = True, + time_context_dim=None, + ff_in=False, + checkpoint=False, + time_depth=1, + attn_mode="softmax", + disable_self_attn=False, + disable_temporal_crossattention=False, + time_mix_legacy: bool = True, + max_time_embed_period: int = 10000, + ): + super().__init__( + in_channels, + n_heads, + d_head, + depth=depth, + dropout=dropout, + attn_type=attn_mode, + use_checkpoint=checkpoint, + context_dim=context_dim, + use_linear=use_linear, + disable_self_attn=disable_self_attn, + ) + self.time_depth = time_depth + self.depth = depth + self.max_time_embed_period = max_time_embed_period + + time_mix_d_head = d_head + n_time_mix_heads = n_heads + + time_mix_inner_dim = int(time_mix_d_head * n_time_mix_heads) + + inner_dim = n_heads * d_head + if use_spatial_context: + time_context_dim = context_dim + + self.time_mix_blocks = nn.ModuleList( + [ + BasicTransformerTimeMixBlock( + inner_dim, + n_time_mix_heads, + time_mix_d_head, + dropout=dropout, + context_dim=time_context_dim, + timesteps=timesteps, + checkpoint=checkpoint, + ff_in=ff_in, + inner_dim=time_mix_inner_dim, + attn_mode=attn_mode, + disable_self_attn=disable_self_attn, + disable_temporal_crossattention=disable_temporal_crossattention, + ) + for _ in range(self.depth) + ] + ) + + assert len(self.time_mix_blocks) == len(self.transformer_blocks) + + self.use_spatial_context = use_spatial_context + self.in_channels = in_channels + + time_embed_dim = self.in_channels * 4 + self.time_mix_time_embed = nn.Sequential( + linear(self.in_channels, time_embed_dim), + nn.SiLU(), + linear(time_embed_dim, self.in_channels), + ) + + self.time_mix_legacy = time_mix_legacy + if self.time_mix_legacy: + if merge_strategy == "fixed": + self.register_buffer("mix_factor", torch.Tensor([merge_factor])) + elif merge_strategy == "learned" or merge_strategy == "learned_with_images": + self.register_parameter( + "mix_factor", torch.nn.Parameter(torch.Tensor([merge_factor])) + ) + elif merge_strategy == "fixed_with_images": + self.mix_factor = None + else: + raise ValueError(f"unknown merge strategy {merge_strategy}") + + self.get_alpha_fn = partial( + get_alpha, + merge_strategy, + self.mix_factor, + apply_sigmoid=apply_sigmoid_to_merge, + is_attn=True, + ) + else: + self.time_mixer = AlphaBlender( + alpha=merge_factor, merge_strategy=merge_strategy + ) + + def forward( + self, + x: torch.Tensor, + context: Optional[torch.Tensor] = None, + # cam: Optional[torch.Tensor] = None, + time_context: Optional[torch.Tensor] = None, + timesteps: Optional[int] = None, + image_only_indicator: Optional[torch.Tensor] = None, + cond_view: Optional[torch.Tensor] = None, + cond_motion: Optional[torch.Tensor] = None, + time_step: Optional[int] = None, + name: Optional[str] = None, + ) -> torch.Tensor: + _, _, h, w = x.shape + x_in = x + spatial_context = None + if exists(context): + spatial_context = context + + if self.use_spatial_context: + assert ( + context.ndim == 3 + ), f"n dims of spatial context should be 3 but are {context.ndim}" + + time_context = context + time_context_first_timestep = time_context[::timesteps] + time_context = repeat( + time_context_first_timestep, "b ... -> (b n) ...", n=h * w + ) + elif time_context is not None and not self.use_spatial_context: + time_context = repeat(time_context, "b ... -> (b n) ...", n=h * w) + if time_context.ndim == 2: + time_context = rearrange(time_context, "b c -> b 1 c") + + x = self.norm(x) + if not self.use_linear: + x = self.proj_in(x) + x = rearrange(x, "b c h w -> b (h w) c") + if self.use_linear: + x = self.proj_in(x) + + if self.time_mix_legacy: + alpha = self.get_alpha_fn(image_only_indicator=image_only_indicator) + + num_frames = torch.arange(timesteps, device=x.device) + num_frames = repeat(num_frames, "t -> b t", b=x.shape[0] // timesteps) + num_frames = rearrange(num_frames, "b t -> (b t)") + t_emb = timestep_embedding( + num_frames, + self.in_channels, + repeat_only=False, + max_period=self.max_time_embed_period, + ) + emb = self.time_mix_time_embed(t_emb) + emb = emb[:, None, :] + + for it_, (block, mix_block) in enumerate( + zip(self.transformer_blocks, self.time_mix_blocks) + ): + # spatial attention + x = block( + x, + context=spatial_context, + time_step=time_step, + name=name + '_' + str(it_) + ) + + x_mix = x + x_mix = x_mix + emb + + # temporal attention + x_mix = mix_block(x_mix, context=time_context, timesteps=timesteps) + if self.time_mix_legacy: + x = alpha.to(x.dtype) * x + (1.0 - alpha).to(x.dtype) * x_mix + else: + x = self.time_mixer( + x_spatial=x, + x_temporal=x_mix, + image_only_indicator=image_only_indicator, + ) + + if self.use_linear: + x = self.proj_out(x) + x = rearrange(x, "b (h w) c -> b c h w", h=h, w=w) + if not self.use_linear: + x = self.proj_out(x) + out = x + x_in + return out + + +class PostHocSpatialTransformerWithTimeMixingAndMotion(SpatialTransformer): + def __init__( + self, + in_channels, + n_heads, + d_head, + depth=1, + dropout=0.0, + use_linear=False, + context_dim=None, + use_spatial_context=False, + timesteps=None, + merge_strategy: str = "fixed", + merge_factor: float = 0.5, + apply_sigmoid_to_merge: bool = True, + time_context_dim=None, + ff_in=False, + checkpoint=False, + time_depth=1, + attn_mode="softmax", + disable_self_attn=False, + disable_temporal_crossattention=False, + time_mix_legacy: bool = True, + max_time_embed_period: int = 10000, + ): + super().__init__( + in_channels, + n_heads, + d_head, + depth=depth, + dropout=dropout, + attn_type=attn_mode, + use_checkpoint=checkpoint, + context_dim=context_dim, + use_linear=use_linear, + disable_self_attn=disable_self_attn, + ) + self.time_depth = time_depth + self.depth = depth + self.max_time_embed_period = max_time_embed_period + + time_mix_d_head = d_head + n_time_mix_heads = n_heads + + time_mix_inner_dim = int(time_mix_d_head * n_time_mix_heads) + + inner_dim = n_heads * d_head + if use_spatial_context: + time_context_dim = context_dim + + camera_context_dim = time_context_dim + motion_context_dim = 4 # time_context_dim + + # Camera attention layer + self.time_mix_blocks = nn.ModuleList( + [ + BasicTransformerTimeMixBlock( + inner_dim, + n_time_mix_heads, + time_mix_d_head, + dropout=dropout, + context_dim=camera_context_dim, + timesteps=timesteps, + checkpoint=checkpoint, + ff_in=ff_in, + inner_dim=time_mix_inner_dim, + attn_mode=attn_mode, + disable_self_attn=disable_self_attn, + disable_temporal_crossattention=disable_temporal_crossattention, + ) + for _ in range(self.depth) + ] + ) + + # Motion attention layer + self.motion_blocks = nn.ModuleList( + [ + BasicTransformerTimeMixBlock( + inner_dim, + n_time_mix_heads, + time_mix_d_head, + dropout=dropout, + context_dim=motion_context_dim, + timesteps=timesteps, + checkpoint=checkpoint, + ff_in=ff_in, + inner_dim=time_mix_inner_dim, + attn_mode=attn_mode, + disable_self_attn=disable_self_attn, + disable_temporal_crossattention=disable_temporal_crossattention, + ) + for _ in range(self.depth) + ] + ) + + assert len(self.time_mix_blocks) == len(self.transformer_blocks) + + self.use_spatial_context = use_spatial_context + self.in_channels = in_channels + + time_embed_dim = self.in_channels * 4 + # Camera view embedding + self.time_mix_time_embed = nn.Sequential( + linear(self.in_channels, time_embed_dim), + nn.SiLU(), + linear(time_embed_dim, self.in_channels), + ) + # Motion time embedding + self.time_mix_motion_embed = nn.Sequential( + linear(self.in_channels, time_embed_dim), + nn.SiLU(), + linear(time_embed_dim, self.in_channels), + ) + + self.time_mix_legacy = time_mix_legacy + if self.time_mix_legacy: + if merge_strategy == "fixed": + self.register_buffer("mix_factor", torch.Tensor([merge_factor])) + elif merge_strategy == "learned" or merge_strategy == "learned_with_images": + self.register_parameter( + "mix_factor", torch.nn.Parameter(torch.Tensor([merge_factor])) + ) + elif merge_strategy == "fixed_with_images": + self.mix_factor = None + else: + raise ValueError(f"unknown merge strategy {merge_strategy}") + + self.get_alpha_fn = partial( + get_alpha, + merge_strategy, + self.mix_factor, + apply_sigmoid=apply_sigmoid_to_merge, + is_attn=True, + ) + else: + self.time_mixer = AlphaBlender( + alpha=merge_factor, merge_strategy=merge_strategy + ) + + def forward( + self, + x: torch.Tensor, + context: Optional[torch.Tensor] = None, + # cam: Optional[torch.Tensor] = None, + time_context: Optional[torch.Tensor] = None, + timesteps: Optional[int] = None, + image_only_indicator: Optional[torch.Tensor] = None, + cond_view: Optional[torch.Tensor] = None, + cond_motion: Optional[torch.Tensor] = None, + time_step: Optional[int] = None, + name: Optional[str] = None, + ) -> torch.Tensor: + _, _, h, w = x.shape + x_in = x + spatial_context = None + if exists(context): + spatial_context = context + + # cond_view: b v 4 h w + # cond_motion: b t 4 h w + b, t, d1 = context.shape # CLIP + v, d2 = cond_view.shape[0]//b, cond_view.shape[1] # VAE + cond_view = torch.nn.functional.interpolate(cond_view, size=(h,w), mode="bilinear") # b*v d h w + spatial_context = context[:,:,None].repeat(1,1,v,1).reshape(-1,1,d1) # (b*t*v) 1 d1 + camera_context = context[:,:,None].repeat(1,1,h*w,1).reshape(-1,1,d1) # (b*t*h*w) 1 d1 + motion_context = cond_view.permute(0,2,3,1).reshape(-1,1,d2) # (b*v*h*w) 1 d2 + + x = self.norm(x) + if not self.use_linear: + x = self.proj_in(x) + x = rearrange(x, "b c h w -> b (h w) c") # 21 x 4096 x 320 + if self.use_linear: + x = self.proj_in(x) + c = x.shape[-1] + + if self.time_mix_legacy: + alpha = self.get_alpha_fn(image_only_indicator=image_only_indicator) + + num_frames = torch.arange(t, device=x.device) + num_frames = repeat(num_frames, "t -> b t", b=b) + num_frames = rearrange(num_frames, "b t -> (b t)") + t_emb = timestep_embedding( + num_frames, + self.in_channels, + repeat_only=False, + max_period=self.max_time_embed_period, + ) + emb_time = self.time_mix_motion_embed(t_emb) + emb_time = emb_time[:, None, :] # b*t x 1 x 320 + + num_views = torch.arange(v, device=x.device) + num_views = repeat(num_views, "t -> b t", b=b) + num_views = rearrange(num_views, "b t -> (b t)") + v_emb = timestep_embedding( + num_views, + self.in_channels, + repeat_only=False, + max_period=self.max_time_embed_period, + ) + emb_view = self.time_mix_time_embed(v_emb) + emb_view = emb_view[:, None, :] # b*v x 1 x 320 + + for it_, (block, time_block, mot_block) in enumerate( + zip(self.transformer_blocks, self.time_mix_blocks, self.motion_blocks) + ): + # Spatial attention + x = block( + x, + context=spatial_context, + ) + + # Camera attention + x = x.view(b, t, v, h*w, c).permute(0,2,1,3,4).reshape(b*v,-1,c) # b*v t*h*w c + x_mix = x + emb_view + x_mix = time_block(x_mix, context=camera_context, timesteps=v) + if self.time_mix_legacy: + x = alpha.to(x.dtype) * x + (1.0 - alpha).to(x.dtype) * x_mix + else: + x = self.time_mixer( + x_spatial=x, + x_temporal=x_mix, + image_only_indicator=image_only_indicator[:,:v], + ) + + # Motion attention + x = x.view(b, v, t, h*w, c).permute(0,2,1,3,4).reshape(b*t,-1,c) # b*t v*h*w c + x_mix = x + emb_time + x_mix = mot_block(x_mix, context=motion_context, timesteps=t) + if self.time_mix_legacy: + x = alpha.to(x.dtype) * x + (1.0 - alpha).to(x.dtype) * x_mix + else: + x = self.time_mixer( + x_spatial=x, + x_temporal=x_mix, + image_only_indicator=image_only_indicator[:,:t], + ) + + x = x.view(b, t, v, h*w, c).reshape(-1,h*w,c) # b*t*v h*w c + + if self.use_linear: + x = self.proj_out(x) + x = rearrange(x, "b (h w) c -> b c h w", h=h, w=w) + if not self.use_linear: + x = self.proj_out(x) + out = x + x_in + return out \ No newline at end of file diff --git a/sgm/modules/video_attention.py b/sgm/modules/video_attention.py new file mode 100644 index 0000000..3d7bd09 --- /dev/null +++ b/sgm/modules/video_attention.py @@ -0,0 +1,302 @@ +import torch + +from ..modules.attention import * +from ..modules.diffusionmodules.util import (AlphaBlender, linear, + timestep_embedding) + + +class TimeMixSequential(nn.Sequential): + def forward(self, x, context=None, timesteps=None): + for layer in self: + x = layer(x, context, timesteps) + + return x + + +class VideoTransformerBlock(nn.Module): + ATTENTION_MODES = { + "softmax": CrossAttention, + "softmax-xformers": MemoryEfficientCrossAttention, + } + + def __init__( + self, + dim, + n_heads, + d_head, + dropout=0.0, + context_dim=None, + gated_ff=True, + checkpoint=True, + timesteps=None, + ff_in=False, + inner_dim=None, + attn_mode="softmax", + disable_self_attn=False, + disable_temporal_crossattention=False, + switch_temporal_ca_to_sa=False, + ): + super().__init__() + + attn_cls = self.ATTENTION_MODES[attn_mode] + + self.ff_in = ff_in or inner_dim is not None + if inner_dim is None: + inner_dim = dim + + assert int(n_heads * d_head) == inner_dim + + self.is_res = inner_dim == dim + + if self.ff_in: + self.norm_in = nn.LayerNorm(dim) + self.ff_in = FeedForward( + dim, dim_out=inner_dim, dropout=dropout, glu=gated_ff + ) + + self.timesteps = timesteps + self.disable_self_attn = disable_self_attn + if self.disable_self_attn: + self.attn1 = attn_cls( + query_dim=inner_dim, + heads=n_heads, + dim_head=d_head, + context_dim=context_dim, + dropout=dropout, + ) # is a cross-attention + else: + self.attn1 = attn_cls( + query_dim=inner_dim, heads=n_heads, dim_head=d_head, dropout=dropout + ) # is a self-attention + + self.ff = FeedForward(inner_dim, dim_out=dim, dropout=dropout, glu=gated_ff) + + if disable_temporal_crossattention: + if switch_temporal_ca_to_sa: + raise ValueError + else: + self.attn2 = None + else: + self.norm2 = nn.LayerNorm(inner_dim) + if switch_temporal_ca_to_sa: + self.attn2 = attn_cls( + query_dim=inner_dim, heads=n_heads, dim_head=d_head, dropout=dropout + ) # is a self-attention + else: + self.attn2 = attn_cls( + query_dim=inner_dim, + context_dim=context_dim, + heads=n_heads, + dim_head=d_head, + dropout=dropout, + ) # is self-attn if context is none + + self.norm1 = nn.LayerNorm(inner_dim) + self.norm3 = nn.LayerNorm(inner_dim) + self.switch_temporal_ca_to_sa = switch_temporal_ca_to_sa + + self.checkpoint = checkpoint + if self.checkpoint: + print(f"{self.__class__.__name__} is using checkpointing") + + def forward( + self, x: torch.Tensor, context: torch.Tensor = None, timesteps: int = None + ) -> torch.Tensor: + if self.checkpoint: + return checkpoint(self._forward, x, context, timesteps) + else: + return self._forward(x, context, timesteps=timesteps) + + def _forward(self, x, context=None, timesteps=None): + assert self.timesteps or timesteps + assert not (self.timesteps and timesteps) or self.timesteps == timesteps + timesteps = self.timesteps or timesteps + B, S, C = x.shape + x = rearrange(x, "(b t) s c -> (b s) t c", t=timesteps) + + if self.ff_in: + x_skip = x + x = self.ff_in(self.norm_in(x)) + if self.is_res: + x += x_skip + + if self.disable_self_attn: + x = self.attn1(self.norm1(x), context=context) + x + else: + x = self.attn1(self.norm1(x)) + x + + if self.attn2 is not None: + if self.switch_temporal_ca_to_sa: + x = self.attn2(self.norm2(x)) + x + else: + x = self.attn2(self.norm2(x), context=context) + x + x_skip = x + x = self.ff(self.norm3(x)) + if self.is_res: + x += x_skip + + x = rearrange( + x, "(b s) t c -> (b t) s c", s=S, b=B // timesteps, c=C, t=timesteps + ) + return x + + def get_last_layer(self): + return self.ff.net[-1].weight + + +class SpatialVideoTransformer(SpatialTransformer): + def __init__( + self, + in_channels, + n_heads, + d_head, + depth=1, + dropout=0.0, + use_linear=False, + context_dim=None, + use_spatial_context=False, + timesteps=None, + merge_strategy: str = "fixed", + merge_factor: float = 0.5, + time_context_dim=None, + ff_in=False, + checkpoint=False, + time_depth=1, + attn_mode="softmax", + disable_self_attn=False, + disable_temporal_crossattention=False, + max_time_embed_period: int = 10000, + ): + super().__init__( + in_channels, + n_heads, + d_head, + depth=depth, + dropout=dropout, + attn_type=attn_mode, + use_checkpoint=checkpoint, + context_dim=context_dim, + use_linear=use_linear, + disable_self_attn=disable_self_attn, + ) + self.time_depth = time_depth + self.depth = depth + self.max_time_embed_period = max_time_embed_period + + time_mix_d_head = d_head + n_time_mix_heads = n_heads + + time_mix_inner_dim = int(time_mix_d_head * n_time_mix_heads) + + inner_dim = n_heads * d_head + if use_spatial_context: + time_context_dim = context_dim + + self.time_stack = nn.ModuleList( + [ + VideoTransformerBlock( + inner_dim, + n_time_mix_heads, + time_mix_d_head, + dropout=dropout, + context_dim=time_context_dim, + timesteps=timesteps, + checkpoint=checkpoint, + ff_in=ff_in, + inner_dim=time_mix_inner_dim, + attn_mode=attn_mode, + disable_self_attn=disable_self_attn, + disable_temporal_crossattention=disable_temporal_crossattention, + ) + for _ in range(self.depth) + ] + ) + + assert len(self.time_stack) == len(self.transformer_blocks) + + self.use_spatial_context = use_spatial_context + self.in_channels = in_channels + + time_embed_dim = self.in_channels * 4 + self.time_pos_embed = nn.Sequential( + linear(self.in_channels, time_embed_dim), + nn.SiLU(), + linear(time_embed_dim, self.in_channels), + ) + + self.time_mixer = AlphaBlender( + alpha=merge_factor, merge_strategy=merge_strategy + ) + + def forward( + self, + x: torch.Tensor, + context: Optional[torch.Tensor] = None, + time_context: Optional[torch.Tensor] = None, + timesteps: Optional[int] = None, + image_only_indicator: Optional[torch.Tensor] = None, + ) -> torch.Tensor: + _, _, h, w = x.shape + x_in = x + spatial_context = None + if exists(context): + spatial_context = context + + if self.use_spatial_context: + assert ( + context.ndim == 3 + ), f"n dims of spatial context should be 3 but are {context.ndim}" + + time_context = context + time_context_first_timestep = time_context[::timesteps] + time_context = repeat( + time_context_first_timestep, "b ... -> (b n) ...", n=h * w + ) + elif time_context is not None and not self.use_spatial_context: + time_context = repeat(time_context, "b ... -> (b n) ...", n=h * w) + if time_context.ndim == 2: + time_context = rearrange(time_context, "b c -> b 1 c") + + x = self.norm(x) + if not self.use_linear: + x = self.proj_in(x) + x = rearrange(x, "b c h w -> b (h w) c") + if self.use_linear: + x = self.proj_in(x) + + num_frames = torch.arange(timesteps, device=x.device) + num_frames = repeat(num_frames, "t -> b t", b=x.shape[0] // timesteps) + num_frames = rearrange(num_frames, "b t -> (b t)") + t_emb = timestep_embedding( + num_frames, + self.in_channels, + repeat_only=False, + max_period=self.max_time_embed_period, + ) + emb = self.time_pos_embed(t_emb) + emb = emb[:, None, :] + + for it_, (block, mix_block) in enumerate( + zip(self.transformer_blocks, self.time_stack) + ): + x = block( + x, + context=spatial_context, + ) + + x_mix = x + x_mix = x_mix + emb + + x_mix = mix_block(x_mix, context=time_context, timesteps=timesteps) + x = self.time_mixer( + x_spatial=x, + x_temporal=x_mix, + image_only_indicator=image_only_indicator, + ) + if self.use_linear: + x = self.proj_out(x) + x = rearrange(x, "b (h w) c -> b c h w", h=h, w=w) + if not self.use_linear: + x = self.proj_out(x) + out = x + x_in + return out diff --git a/sgm/util.py b/sgm/util.py new file mode 100644 index 0000000..66d9b2a --- /dev/null +++ b/sgm/util.py @@ -0,0 +1,275 @@ +import functools +import importlib +import os +from functools import partial +from inspect import isfunction + +import fsspec +import numpy as np +import torch +from PIL import Image, ImageDraw, ImageFont +from safetensors.torch import load_file as load_safetensors + + +def disabled_train(self, mode=True): + """Overwrite model.train with this function to make sure train/eval mode + does not change anymore.""" + return self + + +def get_string_from_tuple(s): + try: + # Check if the string starts and ends with parentheses + if s[0] == "(" and s[-1] == ")": + # Convert the string to a tuple + t = eval(s) + # Check if the type of t is tuple + if type(t) == tuple: + return t[0] + else: + pass + except: + pass + return s + + +def is_power_of_two(n): + """ + chat.openai.com/chat + Return True if n is a power of 2, otherwise return False. + + The function is_power_of_two takes an integer n as input and returns True if n is a power of 2, otherwise it returns False. + The function works by first checking if n is less than or equal to 0. If n is less than or equal to 0, it can't be a power of 2, so the function returns False. + If n is greater than 0, the function checks whether n is a power of 2 by using a bitwise AND operation between n and n-1. If n is a power of 2, then it will have only one bit set to 1 in its binary representation. When we subtract 1 from a power of 2, all the bits to the right of that bit become 1, and the bit itself becomes 0. So, when we perform a bitwise AND between n and n-1, we get 0 if n is a power of 2, and a non-zero value otherwise. + Thus, if the result of the bitwise AND operation is 0, then n is a power of 2 and the function returns True. Otherwise, the function returns False. + + """ + if n <= 0: + return False + return (n & (n - 1)) == 0 + + +def autocast(f, enabled=True): + def do_autocast(*args, **kwargs): + with torch.cuda.amp.autocast( + enabled=enabled, + dtype=torch.get_autocast_gpu_dtype(), + cache_enabled=torch.is_autocast_cache_enabled(), + ): + return f(*args, **kwargs) + + return do_autocast + + +def load_partial_from_config(config): + return partial(get_obj_from_str(config["target"]), **config.get("params", dict())) + + +def log_txt_as_img(wh, xc, size=10): + # wh a tuple of (width, height) + # xc a list of captions to plot + b = len(xc) + txts = list() + for bi in range(b): + txt = Image.new("RGB", wh, color="white") + draw = ImageDraw.Draw(txt) + font = ImageFont.truetype("data/DejaVuSans.ttf", size=size) + nc = int(40 * (wh[0] / 256)) + if isinstance(xc[bi], list): + text_seq = xc[bi][0] + else: + text_seq = xc[bi] + lines = "\n".join( + text_seq[start : start + nc] for start in range(0, len(text_seq), nc) + ) + + try: + draw.text((0, 0), lines, fill="black", font=font) + except UnicodeEncodeError: + print("Cant encode string for logging. Skipping.") + + txt = np.array(txt).transpose(2, 0, 1) / 127.5 - 1.0 + txts.append(txt) + txts = np.stack(txts) + txts = torch.tensor(txts) + return txts + + +def partialclass(cls, *args, **kwargs): + class NewCls(cls): + __init__ = functools.partialmethod(cls.__init__, *args, **kwargs) + + return NewCls + + +def make_path_absolute(path): + fs, p = fsspec.core.url_to_fs(path) + if fs.protocol == "file": + return os.path.abspath(p) + return path + + +def ismap(x): + if not isinstance(x, torch.Tensor): + return False + return (len(x.shape) == 4) and (x.shape[1] > 3) + + +def isimage(x): + if not isinstance(x, torch.Tensor): + return False + return (len(x.shape) == 4) and (x.shape[1] == 3 or x.shape[1] == 1) + + +def isheatmap(x): + if not isinstance(x, torch.Tensor): + return False + + return x.ndim == 2 + + +def isneighbors(x): + if not isinstance(x, torch.Tensor): + return False + return x.ndim == 5 and (x.shape[2] == 3 or x.shape[2] == 1) + + +def exists(x): + return x is not None + + +def expand_dims_like(x, y): + while x.dim() != y.dim(): + x = x.unsqueeze(-1) + return x + + +def default(val, d): + if exists(val): + return val + return d() if isfunction(d) else d + + +def mean_flat(tensor): + """ + https://github.com/openai/guided-diffusion/blob/27c20a8fab9cb472df5d6bdd6c8d11c8f430b924/guided_diffusion/nn.py#L86 + Take the mean over all non-batch dimensions. + """ + return tensor.mean(dim=list(range(1, len(tensor.shape)))) + + +def count_params(model, verbose=False): + total_params = sum(p.numel() for p in model.parameters()) + if verbose: + print(f"{model.__class__.__name__} has {total_params * 1.e-6:.2f} M params.") + return total_params + + +def instantiate_from_config(config): + if not "target" in config: + if config == "__is_first_stage__": + return None + elif config == "__is_unconditional__": + return None + raise KeyError("Expected key `target` to instantiate.") + return get_obj_from_str(config["target"])(**config.get("params", dict())) + + +def get_obj_from_str(string, reload=False, invalidate_cache=True): + module, cls = string.rsplit(".", 1) + if invalidate_cache: + importlib.invalidate_caches() + if reload: + module_imp = importlib.import_module(module) + importlib.reload(module_imp) + return getattr(importlib.import_module(module, package=None), cls) + + +def append_zero(x): + return torch.cat([x, x.new_zeros([1])]) + + +def append_dims(x, target_dims): + """Appends dimensions to the end of a tensor until it has target_dims dimensions.""" + dims_to_append = target_dims - x.ndim + if dims_to_append < 0: + raise ValueError( + f"input has {x.ndim} dims but target_dims is {target_dims}, which is less" + ) + return x[(...,) + (None,) * dims_to_append] + + +def load_model_from_config(config, ckpt, verbose=True, freeze=True): + print(f"Loading model from {ckpt}") + if ckpt.endswith("ckpt"): + pl_sd = torch.load(ckpt, map_location="cpu") + if "global_step" in pl_sd: + print(f"Global Step: {pl_sd['global_step']}") + sd = pl_sd["state_dict"] + elif ckpt.endswith("safetensors"): + sd = load_safetensors(ckpt) + else: + raise NotImplementedError + + model = instantiate_from_config(config.model) + + m, u = model.load_state_dict(sd, strict=False) + + if len(m) > 0 and verbose: + print("missing keys:") + print(m) + if len(u) > 0 and verbose: + print("unexpected keys:") + print(u) + + if freeze: + for param in model.parameters(): + param.requires_grad = False + + model.eval() + return model + + +def get_configs_path() -> str: + """ + Get the `configs` directory. + For a working copy, this is the one in the root of the repository, + but for an installed copy, it's in the `sgm` package (see pyproject.toml). + """ + this_dir = os.path.dirname(__file__) + candidates = ( + os.path.join(this_dir, "configs"), + os.path.join(this_dir, "..", "configs"), + ) + for candidate in candidates: + candidate = os.path.abspath(candidate) + if os.path.isdir(candidate): + return candidate + raise FileNotFoundError(f"Could not find SGM configs in {candidates}") + + +def get_nested_attribute(obj, attribute_path, depth=None, return_key=False): + """ + Will return the result of a recursive get attribute call. + E.g.: + a.b.c + = getattr(getattr(a, "b"), "c") + = get_nested_attribute(a, "b.c") + If any part of the attribute call is an integer x with current obj a, will + try to call a[x] instead of a.x first. + """ + attributes = attribute_path.split(".") + if depth is not None and depth > 0: + attributes = attributes[:depth] + assert len(attributes) > 0, "At least one attribute should be selected" + current_attribute = obj + current_key = None + for level, attribute in enumerate(attributes): + current_key = ".".join(attributes[: level + 1]) + try: + id_ = int(attribute) + current_attribute = current_attribute[id_] + except ValueError: + current_attribute = getattr(current_attribute, attribute) + + return (current_attribute, current_key) if return_key else current_attribute