To train a model on real data run:
python scripts/train.py +experiment=planning.yaml ++data.loaders.train.dataset.percentage_to_use=0.1
To train a model on LidarDM generated data run:
python scripts/train.py +experiment=planning_generated.yaml ++data.loaders.train.dataset.percentage_to_use=0.10
Runs will be named based on a timestamp and stored in the runs/
folder. During training, checkpoints are stored in the checkpoints/
subfolder.
To finetune a model on real data, modify the following command with the PATH_TO_PRETRAINED
with the path to a checkpoint file (ends with .ckpt
) for a planner trained on generated data:
python scripts/train.py +experiment=planning_resume.yaml ++data.loaders.train.dataset.percentage_to_use=0.10 ++model.pretrained=PATH_TO_PRETRAINED
After training a planner model, L2 distance and collision percentage can be evaluated for a checkpoint with:
python scripts/metric/planning_metrics.py +experiment=planning.yaml +pmetrics.trajbank_dir=../../pretrained_models/waymo_trajbank.npy +model.pretrained=PATH_TO_PRETRAINED