RUN.md

Training and Evaluation

nuScenes

After preparing data, run the folloing script to train the model. By default, all models are trained on 8 GPUs.

torchrun --standalone --nnodes=1 --nproc_per_node=8   \
         ./tools/train.py --config-name nusc_det_pp18_aspp_iou_sp \
         data.train_dataset.root_path=/root/to/nuscenes/ \
         dataloader.train.batch_size=6 \
         scheduler.max_lr=0.003 \
         trainer.max_epochs=20 \
         hydra.run.dir=outputs/nusc_pillarnextb

For the reported results, we apply Faded Stratedy, where the copy and paste are remove in the last two epochs. You can add +data.train_dataset.use_gt_sampling=False to disable copy and paste. Currently, we manually stopped the training at epoch 18 and restart the training with the following script:

torchrun --standalone --nnodes=1 --nproc_per_node=8   \
         ./tools/train.py --config-name nusc_det_pp18_aspp_iou_sp \
         data.train_dataset.root_path=/root/to/nuscenes/ \
         dataloader.train.batch_size=6 \
         scheduler.max_lr=0.003 \ 
         trainer.max_epochs=20 \
         hydra.run.dir=outputs/nusc_pillarnextb \
         +data.train_dataset.use_gt_sampling=False \
         +resume_from=epoch_18.pth

Waymo Open Dataset

torchrun --standalone --nnodes=1 --nproc_per_node=8   \
         ./tools/train.py \
         --config-name waymo_det_pp18_aspp_iou_car_sp \
         data.train_dataset.root_path=/path/to/waymo/ \
         dataloader.train.batch_size=3 \
         scheduler.max_lr=0.0015 \
         trainer.max_epochs=36 \
         trainer.eval_every_nepochs=36  \
         hydra.run.dir=outputs/waymo_pillarnextb 

For Waymo, we apply faded strategy in the last 4 epochs.

Note: The results repored in Table 4-6 are trained on 32 GPUs, you can refer to this script for detail. The performance may be slightly different if you are only using 8 GPUs.

For evaluation, please use the official evaluation tools