MolecularDiffusion.modules.models.syncogen.vocab

Vocabulary loading for the SynCoGen port.

Read this before importing anything else from this package.

SynCoGen’s vocabulary is process-global state: constants/constants.py starts with every entry (N_BUILDING_BLOCKS, MAX_ATOMS_PER_BB, COMPATIBILITY, FRAGMENT_ATOMFEATS, …) set to None and load_vocabulary(vocab_dir) fills them in. That is upstream’s own design – train.py calls it at module scope with a ### TODO: Hacky comment, before importing anything else.

The consequence carries over to this port, so it is spelled out here rather than rediscovered: several ported modules bind those constants at import time (BBRxnGraph.VOCAB_NUM_BBS is a class attribute; constants/utils.py does from ...constants import FRAGMENT_ATOMFEATS; SemlaGenerator.__init__ reads N_ATOM_FEATURES). Import any of them before load_vocabulary and they capture None permanently – with no error at import, only a confusing TypeError much later.

So both entry points into this package call ensure_vocabulary first and defer their own imports until after it:

  • modules/models/syncogen/datamodule.py (SyncogenDataModule.__init__)

  • modules/tasks/diffusion_syncogen.py (ModelTaskFactory.build)

ponytail: one vocabulary per process. Loading a second one would leave BBRxnGraph’s class attributes pointing at the first. Fine for train and generate, which each use exactly one; if a caller ever needs two vocabularies side by side, the fix is to move those class attributes onto the instance rather than to add a reload path here.

Attributes

REQUIRED_FILES

Functions

ensure_vocabulary(→ dict[str, Any])

Load vocab_dir into the module-level constants, once per process.

Module Contents

MolecularDiffusion.modules.models.syncogen.vocab.ensure_vocabulary(vocab_dir: str | pathlib.Path) → dict[str, Any]

Load vocab_dir into the module-level constants, once per process.

Idempotent for a repeated call with the same directory. A call with a different directory raises rather than silently half-reloading, because the class attributes captured by the first load would not follow.

MolecularDiffusion.modules.models.syncogen.vocab.REQUIRED_FILES = ('building_blocks.json', 'reactions.json', 'compatibility.pt', 'fragment_features.pt', 'meta.json')