| # compact_v1 construction and reader code | |
| This directory contains the minimum code needed to read the released compact shards and to rebuild the same format from PDB docking poses. | |
| ## Read an existing release | |
| Only `compact_graph_dataset.py` is required for normal training. It provides `CompactGraphDataset`, which opens tensor-only shard files lazily with `torch.load(..., mmap=True)` and reconstructs standard PyTorch Geometric `Data` objects. | |
| ## Build compact shards from PDB poses | |
| Keep these three files in the same directory because the builder imports the other two by filename: | |
| - `build_compact_v1_direct.py` | |
| - `build_graph_unified_enhanced.py` | |
| - `convert_to_compact_v1.py` | |
| The direct builder expects a flat directory with one subdirectory per system containing `protein.pdb`, `ligand.pdb`, and predicted pose `.pdb` files. `materialize_hiqbind_gnncp.py` converts Docking Base common outputs to that layout through hard links. | |
| `build_compact_v1_direct.sbatch` is an example Slurm wrapper only. Review and adapt account, partition, Python environment, paths, CPU count, and memory for your own cluster before use. | |
| Run the bundled unit tests after changing the format code: | |
| ```bash | |
| python -m unittest -v test_compact_graph_dataset.py test_build_compact_v1_direct.py | |
| ``` | |
| See `../../docs/DATASET_USAGE_ZH.md` for the Chinese user guide. | |