Skip to content

train_eval.md

Latest commit

 

History

History
59 lines (38 loc) · 1.61 KB

train_eval.md

File metadata and controls

59 lines (38 loc) · 1.61 KB
Raw

DiffusionDrive Training and Evaluation

1. Cache dataset for faster training and evaluation

# cache dataset for training
bash scripts/caching/caching_training_tran_diff.sh

# cache dataset for evaluation
bash scripts/caching/run_metric_caching.sh

Notes: If you have previously cached the training datasets, you will need to re-cache to properly encode the style features (A key point is to set the agent to diffusiondrive_style_agent or transfuser_style_agent during caching). Otherwise, you may encounter the same issue described here.

2. Training

If your training machine does not have network access, you should download the pretrained ResNet-34 model from huggingface and upload it to your training machine.

DiffusionDrive-Style Model

bash scripts/training/run_diffusiondrive_style_training.sh

Transfuser-Style Model

bash scripts/training/run_transfuser_style_training.sh

AD-MLP-Style Model

bash scripts/training/run_ego_mlp_style_agent_training.sh

3. Evaluation

All the ckpts are open-sourced in Huggingface.

You can modify the ckpt's path in the following scripts and run the evaluation as follows:

DiffusionDrive-Style Model

bash scripts/evaluation/run_diffusiondrive_style.sh

Transfuser-Style Model

bash scripts/evaluation/run_transfuser_style.sh

AD-MLP-Style Model

bash scripts/evaluation/run_ego_mlp_style_agent_pdm_score_evaluation.sh