Training and Evaluation

Training

Download the ResNet-50 image-backbone initialization. A processed htc_r50_backbone.pth is available here.

Train with 4 GPUs:

bash tools/dist_train.sh PATH_TO_CONFIG 4 \
  --run-dir CHECKPOINT_SAVE_DIR \
  --model.encoders.camera.backbone.init_cfg.checkpoint PATH_TO_PRETRAIN

Train with 1 GPU:

python tools/train.py PATH_TO_CONFIG \
  --run-dir CHECKPOINT_SAVE_DIR \
  --model.encoders.camera.backbone.init_cfg.checkpoint PATH_TO_PRETRAIN

PATH_TO_PRETRAIN may be the processed ResNet-50 checkpoint described above. Training writes the resolved configuration, logs, and checkpoints to CHECKPOINT_SAVE_DIR.

Evaluation

Evaluate with 4 GPUs:

bash tools/dist_test.sh PATH_TO_CONFIG PATH_TO_WEIGHT 4

Evaluate with 1 GPU:

python tools/test.py PATH_TO_CONFIG PATH_TO_WEIGHT

The evaluation protocol and masks are controlled by the selected configuration. In particular, use the wo_mask Occ3D-nuScenes configuration when reporting RayIoU without camera masks by adding:

evaluation:
  occ_metric: Ray-IoU

Inference benchmark

Measure end-to-end inference throughput on one GPU with:

python tools/benchmark.py PATH_TO_CONFIG \
  --checkpoint PATH_TO_WEIGHT \
  --samples 50