MolecularDiffusion.modules.models.chefnmr.sidecar¶

Row-aligned sidecar arrays for ChefNMR.

The platform’s pointcloud batch has a per-atom array channel (node_features) and a per-molecule scalar channel (target_fields), but no per-molecule array channel. ChefNMR needs two of those: a (10080,) binned NMR condition and a (C, N, 3) ground-truth conformer stack. Both ride memmapped .npy files keyed by row index and are joined inside the task, off batch["xyz"] – the same pattern diffusion_diffsmol.py uses for its shape latent, and the reason the data layer needs no change.

Why memmap and not one .pt dict like DiffSMol’s: the condition is 40 kB per molecule (27x DiffSMol’s latent), and torch.load would pull the whole map into RAM. np.load(..., mmap_mode="r")[i] is O(1) resident and reads 40 kB per item.

Layout, all written in one pass by docs/model_integrations/chefnmr/scripts/convert_dataset.py:

<prefix>.db            ASE db; row i <-> xyz == f"db_entry_{i}"
<prefix>_cond.npy      (R, h_dim + c_dim) float32
<prefix>_conf.npy      (R, max_C, max_n_atoms, 3) float32, zero-padded
<prefix>_nconf.npy     (R,) int32 -- real conformers per row
<prefix>_meta.json     R, max_C, max_n_atoms, atom_decoder, sigma_data,
                       db sha256 + path, split, sparsity report

A miss is a hard error, not a fallback. DiffSMol falls back to a zero latent on a cache miss; here a zero condition is the classifier-free unconditional branch, so the model would happily emit a plausible molecule of the right formula that has nothing to do with the spectrum and the run would look fine. Every lookup failure raises.

Attributes¶

logger

Classes¶

ChefNMRSidecar

The three row-aligned arrays plus the conversion metadata.

Functions¶

`load_sidecar`(→ Optional[ChefNMRSidecar])	Open the memmaps and cross-check them against `_meta.json`.
`parse_row_index`(→ int)	`"db_entry_17"` -> `17`, with a message naming the real cause.
`sha256_file`(→ str)	Streamed sha256 of a file.

Module Contents¶

class MolecularDiffusion.modules.models.chefnmr.sidecar.ChefNMRSidecar¶

The three row-aligned arrays plus the conversion metadata.

cond: numpy.ndarray¶

property cond_dim: int¶

conf: numpy.ndarray¶

property max_n_atoms: int¶

meta: dict¶

n_conf: numpy.ndarray¶

property n_rows: int¶

MolecularDiffusion.modules.models.chefnmr.sidecar.load_sidecar(cond_path: str | None, conf_path: str | None, meta_path: str | None) → ChefNMRSidecar | None¶

Open the memmaps and cross-check them against _meta.json.

Returns None when no paths are configured – which is the generate-from-checkpoint case, where the corpus comes from the interference config instead and the task never trains.

MolecularDiffusion.modules.models.chefnmr.sidecar.parse_row_index(xyz: object, n_rows: int) → int¶

"db_entry_17" -> 17, with a message naming the real cause.

PointCloudDataset.save_pickle(cheap_data=True) nulls self.xyzs, which destroys the join key outright; so does pointing the task at a db that was not the one the sidecar was built from.

MolecularDiffusion.modules.models.chefnmr.sidecar.sha256_file(path: str, chunk: int = 8 << 20) → str¶: Streamed sha256 of a file.

MolecularDiffusion.modules.models.chefnmr.sidecar.logger¶