| # MagNET inference dependencies (CPU; tested on Apple Silicon arm64 and Linux x86-64). |
| # |
| # Two non-obvious points are baked in below: |
| # 1. torch_scatter and torch_cluster are NOT on plain PyPI for macOS arm64. They come |
| # from the PyG wheel index, which is keyed to the EXACT torch version. The |
| # --find-links line points pip at the matching prebuilt wheels for torch 2.5.0 + CPU. |
| # 2. torch_geometric is PINNED to 2.6.x. From 2.7 onward, radius_graph requires pyg-lib |
| # (which has no Apple-Silicon wheel); 2.6.x still dispatches radius_graph to |
| # torch_cluster, which is what the model needs. |
| # |
| # Install torch first, then this file, or just run this file twice if a fresh |
| # environment resolves torch_scatter before torch: |
| # pip install torch==2.5.0 |
| # pip install -r requirements.txt |
| # |
| # On a Linux machine with CUDA, `pip install torch==2.5.0` defaults to the multi-gigabyte GPU build. |
| # This is a CPU stack, so install the CPU build explicitly first: |
| # pip install torch==2.5.0 --index-url https://download.pytorch.org/whl/cpu |
|
|
| --find-links https://data.pyg.org/whl/torch-2.5.0+cpu.html |
|
|
| torch==2.5.0 |
| torch_scatter |
| torch_cluster |
| torch_geometric==2.6.1 |
| pytorch_lightning>=2.0 |
| e3nn>=0.5 |
| numpy |
|
|
| # Only needed to run the bundled examples/tests that read the .hdf5 datasets |
| # (the model itself does not import h5py): |
| h5py |
|
|
