train_and_eval.md

Getting started with OccFormer

For most of our experiments, we train the model with 8 RTX 3090 GPUs with 24G memory. Unfortunately, we already use the pytorch checkpoint (for all models) and fp16 (only for R101-DCN experiment on nuScenes) to reduce the memory. Therefore, you may need similar hardwares to reproduce the training results.

Before start training, download the corresponding pretrained backbones from the release page and put them under the folder ckpts/. The weights include EfficientNetB7 for KITTI and R50 & R101-DCN for nuScenes.

Training

bash tools/dist_train.sh $CONFIG 8

During the training process, the model is evaluated on the validation set after every epoch. The checkpoint with best performance will be saved. The output logs and checkpoints will be available at work_dirs/$CONFIG.

Evaluation

Evaluate with 1 GPU:

python tools/test.py $YOUR_CONFIG $YOUR_CKPT --eval=bbox

The single-GPU inference will print the current performance after each iteration, which can serve as a quick indicator.

Evaluate with 8 GPUs:

bash tools/dist_test.sh $YOUR_CONFIG $YOUR_CKPT 8