Skip to content

Latest commit

 

History

History
55 lines (36 loc) · 1.97 KB

TRAIN.md

File metadata and controls

55 lines (36 loc) · 1.97 KB
Raw

Training and Testing guide

Training

1) ScanNetV2 dataset

Pretrain the 3D Unet backbone from scratch

python3 tools/train.py configs/scannetv2/isbnet_backbone_scannetv2.yaml --only_backbone --exp_name pretrain_backbone

We also provided the pre-trained 3D backbone models of ScanNetV2 val, ScanNetV2-200 val, S3DIS val Area 5, and STPLS3D val. Alternatively, we can also finetune other pre-trained backbones from SoftGroup or SSTNet with fewer epoches by set the pretrain path in the config file to the corresponding pre-trained weights.

Train our ISBNet

python3 tools/train.py configs/scannetv2/isbnet_scannetv2.yaml --trainall --exp_name default

By default, we set batch_size=12 on a single V100 GPU.

2) S3DIS dataset

# Pretrain step
python3 tools/train.py configs/s3dis/isbnet_backbone_s3dis_area5.yaml --only_backbone  --exp_name default

# Train entire model
python3 tools/train.py configs/s3dis/isbnet_s3dis_area5.yaml --trainall  --exp_name default

3) STPLS3D dataset

# Pretrain step
python3 tools/train.py configs/stpls3d/isbnet_backbone_stpls3d.yaml --only_backbone  --exp_name default

# Train entire model
python3 tools/train.py configs/stpls3d/isbnet_stpls3d.yaml --trainall  --exp_name default

Inference

1) For evaluation (on ScanNetV2 val, S3DIS, and STPLS3D)

python3 tools/test.py configs/<config_file> <checkpoint_file>

2) For exporting predictions (i.e., to submit results to ScanNetV2 hidden benchmark)

python3 tools/test.py configs/<config_file> <checkpoint_file> --out <output_dir>