No public description

PiperOrigin-RevId: 684582762

Author: tensorflower-gardener

official/projects/waste_identification_ml/README.md

@@ -75,6 +75,69 @@ Material Form Model V2| MobileNet | saved model | [click here](https://storage.g
     --config_file="config.yaml"`
 10. You can also start a screen session and run the training in the background.
 
+## Config file parameters
+
+-   `annotation_file` - path to the validation file in COCO JSON format.
+-   `init_checkpoint` - path to the checkpoints for transfer learning.
+-   `init_checkpoint_modules` - to load both the backbone or decoder or any one
+     of them.
+-   `freeze_backbone` - if you want to freeze your backbone or not while
+     training.
+-   `input_size` - image size according to which the model is trained.
+-    `num_classes` - total number of classes + 1 ( background )
+-   `per_category_metrics` - in case you need metric for each class
+-   `global_batch_size` - batch size.
+-   `input_path` - path to the dataset set.
+-   `parser` - contains the data augmentation operations.
+-   `steps_per_loop` - number of steps to complete one epoch. It's usually
+     `training tal data size / batch size`.
+-   `summary_interval` - how often you want to plot the metric
+-   `train_steps` - total steps for training. Its equal to
+     `steps_per_loop x epochs`
+-   `validation_interval` - how often do you want to evaluate the validation
+     data.
+-   `validation_steps` - steps to cover validation data. Its equal to
+     `validation data size / batch size`
+-   `warmup_learning_rate` - it is a strategy that gradually increases the
+     learning rate from a very low value to a desired initial learning rate over
+     a predefined number of iterations or epochs.
+     To stabilize training in the early stages by allowing the model to adapt to
+     the data slowly before using a higher learning rate.
+-   `warmup_steps` - steps for the warmup learning rate
+-   `initial_learning_rate` - The initial learning rate is the value of the
+     learning rate at the very start of the training process.
+-   `checkpoint_interval` - number of steps to export the model.
+
+A common practice to calculate the parameters are below:
+
+`total_training_samples = 4389
+total_validation_samples = 485
+
+train_batch_size = 512
+val_batch_size = 128
+num_epochs = 700
+warmup_learning_rate = 0.0001
+initial_learning_rate = 0.001
+
+steps_per_loop = total_training_samples // train_batch_size
+summary_interval = steps_per_loop
+train_steps = num_epochs * steps_per_loop
+validation_interval = steps_per_loop
+validation_steps = total_validation_samples // val_batch_size
+warmup_steps = steps_per_loop * 10
+checkpoint_interval = steps_per_loop * 5
+decay_steps = int(train_steps)
+
+print(f'steps_per_loop: {steps_per_loop}')
+print(f'summary_interval: {summary_interval}')
+print(f'train_steps: {train_steps}')
+print(f'validation_interval: {validation_interval}')
+print(f'validation_steps: {validation_steps}')
+print(f'warmup_steps: {warmup_steps}')
+print(f'warmup_learning_rate: {warmup_learning_rate}')
+print(f'initial_learning_rate: {initial_learning_rate}')
+print(f'decay_steps: {decay_steps}')
+print(f'checkpoint_interval: {checkpoint_interval}')`
 
 ## Authors and Maintainers
 - Umair Sabir