Training

All our experiments use 4 machines with 8 RTX 4090 GPUs. We recommend using a large batch size and a multi-machine training strategy.
If using a single machine with 8 RTX 4090 GPUs or a similar configuration, we suggest freezing the perception module during Step 2 and Step 3. This can achieve similar performance while reducing training time.

Step 1: Train the Perception Module

First, download the pretrained V99 backbone to data. Then, train the perception module separately before training other modules.

• Set ONLY_PERCEPTION to True.
• Specify the path to the V99_PRETRAINED_PATH.
• The model output will be saved in the experiment log.

sh scripts/training/run_goalflow_training_perception.sh

Step 2: Train the Trajectory Planning Module

• Set ONLY_PERCEPTION to False.
• To accelerate training, you can set FREEZE_PERCEPTION to True.(Optional)
• Specify the path to the perception model using CHECKPOINT_PATH and FEATURE_PATH obtrained from run_dataset_cache_trainval.sh.

sh scripts/training/run_goalflow_training_traj.sh

Step 3: Train the Goal Point Construction Module (Optional)

• To accelerate training, you can set FREEZE_PERCEPTION to True.(Optional)
• Specify the path to the imported model using CHECKPOINT_PATH and FEATURE_PATH obtrained from run_dataset_cache_trainval.sh.

sh scripts/training/run_goalflow_training_navi.sh