MolecularDiffusion.assets

Resolve zoo assets (pretrained weights, datasets) by symbolic name.

A config says ${asset:kgdiff/pretrained} instead of a machine-specific path, and this module turns that into a real directory under the local cache. The rest of the platform never learns the zoo exists: the resolver hands back a plain path string, so cli/generate.py’s file-or-directory handling, its edm_chem.pkl / edm_stat.pkl sidecar lookups and its ../../config.yaml walk all keep working unchanged.

Three properties worth stating, because the design leans on them:

  • An asset is a directory. Every file listed under one assets: entry lands in the same directory, which is what preserves the sidecar co-location contract that cli/generate.py depends on.

  • Literal paths still work. ${asset:...} is opt-in syntax. A config carrying an ordinary path never reaches this module, so every existing config behaves exactly as before.

  • Resolution never touches the network. A cache hit is a Path.exists() call. Fetching is explicit (MolCraftDiff zoo fetch), because a config typo must not silently start a multi-gigabyte download inside a Hydra compose.

The manifest lives in MANIFEST (zoo.yaml, shipped in the wheel).

Attributes

MANIFEST

Functions

assets_root(→ pathlib.Path)

Where fetched assets live. $MOLCRAFT_ASSETS wins if set.

entry(→ dict[str, Any])

The manifest entry for name, with aliases already resolved.

fetch(→ pathlib.Path)

Download an asset into the cache and verify every file's sha256.

local_path(→ pathlib.Path)

Absolute cache path for name.

manifest(→ dict[str, Any])

The parsed zoo.yaml, read once per process.

register(→ None)

Register the ${asset:...} OmegaConf resolver.

resolve(→ str)

The ${asset:...} resolver body. Returns an absolute path string.

verify(→ list[tuple[str, str]])

Re-hash an asset's files. Returns [(relpath, status), ...].

Module Contents

MolecularDiffusion.assets.assets_root() → pathlib.Path

Where fetched assets live. $MOLCRAFT_ASSETS wins if set.

MolecularDiffusion.assets.entry(name: str) → dict[str, Any]

The manifest entry for name, with aliases already resolved.

MolecularDiffusion.assets.fetch(name: str, force: bool = False) → pathlib.Path

Download an asset into the cache and verify every file’s sha256.

Re-fetching is a no-op: a file already present with the right hash is skipped. Pass force=True to redownload regardless.

MolecularDiffusion.assets.local_path(name: str) → pathlib.Path

Absolute cache path for name.

Aliases resolve before the path is built, so ipdiff/data and kgdiff/data are literally the same directory – shared corpora are stored once, never copied per model.

MolecularDiffusion.assets.manifest() → dict[str, Any]

The parsed zoo.yaml, read once per process.

MolecularDiffusion.assets.register() → None

Register the ${asset:...} OmegaConf resolver.

Called from MolecularDiffusion/__init__.py so it covers both the CLI (which composes via cli/_hydra.py) and plain library use (OmegaConf.load in a notebook, or cli/generate.py’s own load of a checkpoint’s training config).

MolecularDiffusion.assets.resolve(name: str) → str

The ${asset:...} resolver body. Returns an absolute path string.

Raises FileNotFoundError naming the exact fetch command when the asset is absent, rather than auto-downloading – unless MOLCRAFT_ASSETS_AUTOFETCH=1 is set.

MolecularDiffusion.assets.verify(name: str) → list[tuple[str, str]]

Re-hash an asset’s files. Returns [(relpath, status), ...].

Status is ok, missing, sha mismatch, or no sha in manifest. This is what makes a locally-built, non-redistributable asset trustworthy: the manifest ships the hash even when it cannot ship the bytes.

MolecularDiffusion.assets.MANIFEST