Model Weights¶

All trained weights are hosted on Hugging Face:

https://huggingface.co/Membrizard/ml_conformer_generator

Available Weights¶

File	Format	Model	Notes
`edm_moi_chembl_15_39.pt`	PyTorch	EDM (core, 24M params)	Default for `MLConformerGenerator`
`adj_mat_seer_chembl_15_39.pt`	PyTorch	AdjMatSeer (core, 22M params)	Default for `MLConformerGenerator`
`small_edm_moi_chembl_15_39.pt`	PyTorch	EDM (distilled, 8.9M params)	Lower memory, faster
`small_adj_mat_seer_obabel_15_39.pt`	PyTorch	AdjMatSeer (distilled, 1.6M params)	Fine-tuned to match OpenBabel bond perception; higher validity
`edm_moi_chembl_6_39_fragments.pt`	PyTorch	EDM (fragments, 6–39 atoms)	Generator model for Inertial Fragment Matching
`edm_moi_chembl_15_39_inpaint.pt`	PyTorch	EDM (inpainting variant)	Alternative weights for fixed-fragment generation
`egnn_chembl_15_39.onnx`	ONNX	EGNN denoiser (core)	Default for `MLConformerGeneratorONNX` and the JS package
`adj_mat_seer_chembl_15_39.onnx`	ONNX	AdjMatSeer (core)	Default for `MLConformerGeneratorONNX` and the JS package
`small_egnn_chembl_15_39.onnx`	ONNX	EGNN denoiser (distilled)
`small_adj_mat_seer_obabel_15_39.onnx`	ONNX	AdjMatSeer (distilled)

The PyTorch generator infers the hidden width and base number of diffusion timesteps from the checkpoint, so full and distilled weights are interchangeable — only the file names differ.

Python: Automatic Resolution¶

Both MLConformerGenerator and MLConformerGeneratorONNX resolve weight files through a built-in WeightsManager. For every weights argument the following order applies:

If the value is an existing local path, it is used as is.
Otherwise the file is looked up in the local cache (~/.mlconfgen_weights by default).
If not cached, the file is downloaded from the Hugging Face repository into the cache. huggingface_hub is installed automatically if missing.

from mlconfgen import MLConformerGenerator

# Default core weights, downloaded on first run
model = MLConformerGenerator()

# Distilled weights, by name
model = MLConformerGenerator(
    edm_weights="small_edm_moi_chembl_15_39.pt",
    adj_mat_seer_weights="small_adj_mat_seer_obabel_15_39.pt",
)

# Local files
model = MLConformerGenerator(
    edm_weights="./checkpoints/my_edm.pt",
    adj_mat_seer_weights="./checkpoints/my_seer.pt",
)

Downloads are protected by a lock file so that several processes starting concurrently do not download the same file twice.

Listing and clearing¶

model.list_weights()
# {'remote': ['adj_mat_seer_chembl_15_39.pt', 'edm_moi_chembl_15_39.pt', ...],
#  'local':  ['edm_moi_chembl_15_39.pt', 'adj_mat_seer_chembl_15_39.pt']}

model.clear_cache()   # removes everything under the cache directory

list_weights() filters by format — .pt for the PyTorch generator, .onnx for the ONNX generator.

Custom cache directory¶

from mlconfgen.utils import WeightsManager

wm = WeightsManager(cache_dir="/shared/mlconfgen_weights")
path = wm.resolve("edm_moi_chembl_15_39.pt")           # -> Path to the file
path = wm.resolve("edm_moi_chembl_15_39.pt", force_download=True)

JavaScript: Local Files¶

The npm package does not download weights. Download the ONNX files from Hugging Face and pass them to createGenerator as file paths (Node) or Uint8Array buffers (browser):

const gen = await createGenerator({
  ort,
  egnnOnnx: "./egnn_chembl_15_39.onnx",
  adjMatSeerOnnx: "./adj_mat_seer_chembl_15_39.onnx",
});

See Browser Usage for loading weights from a user's machine without re-hosting them.