spacr.classification_pixels

Versioned crop decoding shared by classification and its checkpoints.

Functions

checkpoint_policy(metadata, *[, announce])	Read a checkpoint's decoder contract, preserving historical untagged models.
initialization_policy(path)	Select decoding before training loaders are built for a resumed model.
loader_policy(loader)	Find the decoding record through loaders, subsets and combined datasets.
read_classification_image(source[, policy, fmt, ...])	Decode a path or archive stream using the model's recorded policy.
training_preprocessing(train_loader, validation_loader)	Record actual loader decoding and reject checkpoint or validation conflicts.
validate_policy(policy)	Accept only implemented crop decoding policies; never guess a new one.

Module Contents

spacr.classification_pixels.checkpoint_policy(metadata, *, announce=False)

Read a checkpoint’s decoder contract, preserving historical untagged models.

Parameters:

• metadata – artifact mapping returned by load_model_artifact.

• announce – explain the legacy fallback when the contract is absent.

Returns:

validated policy name. Missing metadata selects stored PIL RGB.

spacr.classification_pixels.initialization_policy(path)

Select decoding before training loaders are built for a resumed model.

Parameters:

path – trusted checkpoint path, or None for a new training run.

Returns:

saved policy, legacy fallback, or declared uint8 for new training.

spacr.classification_pixels.loader_policy(loader)

Find the decoding record through loaders, subsets and combined datasets.

Unknown external tensor datasets return None. Mixed recorded policies are refused because a single checkpoint cannot describe both transformations.

Parameters:

loader – loader or dataset exposing crop_loading_policy, dataset or datasets.

Returns:

a validated policy name, or None if no policy is recorded.

Raises:

ValueError – a policy is unsupported or combined datasets disagree, including a mixture of recorded and unknown policies.

spacr.classification_pixels.read_classification_image(source, policy=DECLARED_UINT8, *, fmt=None, legacy_orient=False)

Decode a path or archive stream using the model’s recorded policy.

Parameters:

• source – image path or seekable binary stream, as accepted by PIL.

• policy – declared uint8 decoding or historical stored PIL RGB.

• fmt – source format; None resolves the path’s per-file crop marker. Archive callers must supply a format (1 when unmarked).

• legacy_orient – retain the labelled loader’s historical EXIF handling.

Returns:

independently owned PIL RGB image with its file handle closed.

spacr.classification_pixels.training_preprocessing(train_loader, validation_loader, preprocessing=None, checkpoint=None)

Record actual loader decoding and reject checkpoint or validation conflicts.

Parameters:

• train_loader – training loader or dataset whose decoding is inspected.

• validation_loader – validation loader or dataset checked for agreement.

• preprocessing – optional preprocessing mapping copied into the result; an explicit policy must agree with a recorded training policy.

• checkpoint – optional loaded checkpoint metadata checked against the selected policy; an untagged checkpoint uses the historical policy.

Returns:

a new preprocessing dictionary, including the policy when known. Unknown external datasets are not assigned a decoder automatically.

Raises:

ValueError – a supplied policy is unsupported or the recorded training, validation, preprocessing or checkpoint policies conflict.

spacr.classification_pixels.validate_policy(policy)

Accept only implemented crop decoding policies; never guess a new one.

Parameters:

policy – declared_uint8_v1 or stored_pil_v1.

Returns:

the unchanged, validated policy name.

Raises:

ValueError – the supplied policy is not implemented.