Skip to content

TRAIN_EVAL.md

Latest commit

 

History

History
97 lines (67 loc) · 4.28 KB

TRAIN_EVAL.md

File metadata and controls

97 lines (67 loc) · 4.28 KB

Train/Eval Models

Evaluation Example

Please make sure you have prepared the environment and the nuScenes dataset. You can check it by simply evaluating the pre-trained first-stage(track_map) model as follows:

cd UniAD
./tools/uniad_dist_eval.sh ./projects/configs/stage1_track_map/base_track_map.py ./ckpts/uniad_base_track_map.pth 8 # (8 is the number of GPUs)

# For slurm users:
# ./tools/uniad_slurm_eval.sh YOUR_PARTITION ./projects/configs/stage1_track_map/base_track_map.py ./ckpts/uniad_base_track_map.pth 8

If everything is prepared properly, the output results should be:

Aggregated results: 
AMOTA	0.394 
AMOTP	1.316
RECALL	0.484

Note: If you evaluate with different number of GPUs rather than 8, the results might be slightly different.

GPU Requirements

How to run UniAD lightweight, say just two 3090s at an affordable cost, is definitely an incoming feature. Stay tuned!

UniAD is trained in two stages. The first stage is to train the perception modules (e.g., track and map), and the second stage initializes the weights trained from last stage and optimizes all task modules together. It's recommended to use at least 8 GPUs for training in both two stages. Training with fewer GPUs is fine but would cost more time.

The first-stage training takes ~ 50 GB GPU memory, ~ 2 days for 6 epochs on 8 A100 GPUs.

  • HINT: To save GPU memory, you can change queue_length=5 to 3 which will slightly degrade the tracking performance. Then the training would take ~30 GB GPU memory and is affordable for V100 GPUs (32GB version).

The second-stage training takes ~ 17 GB GPU memory, ~ 4 days for 20 epochs on 8 A100 GPUs.

  • NOTE: Compared to the first-stage, much less GPU memory is required because we freeze the BEV encoder in this stage to focus on learning task-specific queries. Due to this, you can run the second-stage training on V100 or 3090 devices.

Train

Bevformer Backbone Training

we integrated BEVFormer training (you also can use other BEV backbone) into the UniAD code repository in this version.

# Training
./tools/uniad_dist_train.sh ./projects/configs/bevformer/bevformer_base.py N_GPUS

# Evaluation
./tools/uniad_dist_train.sh ./projects/configs/bevformer/bevformer_base.py UniAD/ckpts/bevformer_r101_dcn_24ep.pth N_GPUS

Training Command

# N_GPUS is the number of GPUs used. Recommended >=8.
./tools/uniad_dist_train.sh ./projects/configs/stage1_track_map/base_track_map.py N_GPUS

# For slurm users:
# ./tools/uniad_slurm_train.sh YOUR_PARTITION ./projects/configs/stage1_track_map/base_track_map.py N_GPUS

Evaluation

Eval Command

# N_GPUS is the number of GPUs used.  Recommended =8.
# If you evaluate with different number of GPUs rather than 8, the results might be slightly different.

./tools/uniad_dist_eval.sh ./projects/configs/stage1_track_map/base_track_map.py /PATH/TO/YOUR/CKPT.pth N_GPUS

# For slurm users:
# ./tools/uniad_slurm_eval.sh YOUR_PARTITION ./projects/configs/stage1_track_map/base_track_map.py /PATH/TO/YOUR/CKPT.pth N_GPUS

Visualization

visualization Command

# please refer to  ./tools/uniad_vis_result.sh
python ./tools/analysis_tools/visualize/run.py \
    --predroot /PATH/TO/YOUR/RESULTS.pkl \
    --out_folder /PATH/TO/YOUR/OUTPUT \
    --demo_video test_demo.avi \
    --project_to_cam True

<- Last Page: Prepare The Dataset