export_weights

export_weights(model, spec, path, *, extra=None)

Write a model’s weights as safetensors, with the sidecar that describes them.

The sidecar is not optional here. Weights without a record of what they are for are the thing this module exists to prevent, and the moment to write it is while the information is still at hand.

Parameters

Name	Type	Description	Default
model		The trained network.	required
spec		The model’s specification, which is what the sidecar records.	required
path	str | Path	Where to write. The sidecar goes beside it.	required
extra	dict | None	Anything else worth recording - for a detector, the anchors, which are not optional metadata: the same weights read with different anchors decode every box scaled by a fixed factor, and nothing in the output says so.	None

Returns

Name	Type	Description
	Path	The path written.

The sidecar records the architecture, the shape, the channels, the classes, the blocks, the filters and trained_with - the engine version that produced the file. Loading refuses a model the weights do not belong to before touching torch, because a wrong channel count fails in load_state_dict anyway while weights trained on a different colormap with the same class count load cleanly and predict nonsense.

safetensors rather than a pickle, which would make a weights file a program somebody else wrote.