Skip to content

training.md

Latest commit

 

History

History
76 lines (58 loc) · 2.22 KB

training.md

File metadata and controls

76 lines (58 loc) · 2.22 KB
Raw

🏋️ Training

We provide preprocessed datasets and training scripts to reproduce the results in the paper.

📦 Dataset Preparation

Download the dataset from Google Drive and place it under the following directory structure:

{project_root}/
├── src/
├── data/
│ └── abc_processed/
│ ├── v1_1_grid8.h5
│ ├── deepcad_30_train.txt
│ ├── deepcad_30_val.txt
│ ├── deepcad_30_test.txt
│ ├── deepcad_30_pkl_absence.txt
│ ├── abc_50_train.txt
│ ├── abc_50_val.txt
│ ├── abc_50_test.txt
│ └── abc_50_pkl_absence.txt
└── ...
  • deepcad_30_* contains the train/val/test splits for the DeepCAD dataset.
  • abc_50_* contains the splits for the ABC 50-face subset.
  • The *_pkl_absence.txt files list missing samples relative to BrepGen. For the DeepCAD dataset, this should be empty, so do not worry about it.

🚀 Training

We recommend the following hardware configurations:

🔹 DeepCAD dataset (2 × 24GB GPUs)

python scripts/run.py logs/brepdiff/deepcad --config-path configs/deepcad.yaml

Expected training time: ~4 days with 2×NVIDIA 3090 (24GB) GPUs.

🔹 ABC 50-face dataset (4 × 48GB GPUs)

python scripts/run.py logs/brepdiff/abc --config-path configs/abc.yaml

Expected training time: ~7 days with 4×NVIDIA A6000 (48GB) GPUs.

Add the --debug flag to quickly verify that training runs correctly. By default, Weights & Biases (W&B) logging is enabled. Use the --wandb-offline flag to disable W&B logging.

🧯 Troubleshooting

❌ ValueError: current limit exceeds maximum limit

You may see an error like this:

Traceback (most recent call last):
  File "scripts/run.py", line 18, in <module>
    resource.setrlimit(resource.RLIMIT_NOFILE, (8192, rlimit[1]))
ValueError: current limit exceeds maximum limit

This means your system’s file descriptor soft limit is too high. To resolve it, lower the value in scripts/run.py:

# Before:
resource.setrlimit(resource.RLIMIT_NOFILE, (8192, rlimit[1]))

# After:
resource.setrlimit(resource.RLIMIT_NOFILE, (1024, rlimit[1]))