blinklinmult.paper.session

Running a frozen ONNX graph, with numpy and onnxruntime alone.

No torch here, and none reachable from here. That is the whole point of shipping the 1.x models as graphs: their inference path cannot be broken by a PyTorch upgrade, a linmult release, or anything else this repository does later.

The class is deliberately thin. It holds a session, remembers what the graph’s inputs are called, and returns numpy arrays – BlinkDetector supplies the normalisation and the event extraction on top, exactly as it does for the PyTorch-backed v2 model.

Attributes

logger	Module-level logger.
IMAGENET_MEAN	Per-channel mean the 1.x models were trained against.
IMAGENET_STD	Per-channel standard deviation. See IMAGENET_MEAN.
CROPS_INPUT	Name of the eye-crop input in every exported graph.
FEATURES_INPUT	Name of the descriptor input, present only in the two-stream graph.
SEQUENCE_OUTPUT	Name of the per-frame output in every exported graph.
CLIP_OUTPUT	Name of the pooled output, present only in the two-stream graph.

Exceptions

SessionError

Raised when a graph cannot be loaded or is fed the wrong inputs.

Classes

OnnxModel

A frozen 1.x model, backed by an onnxruntime session.

Functions

load_onnx(→ OnnxModel)

Open an exported graph.

Module Contents

blinklinmult.paper.session.logger

Module-level logger.

blinklinmult.paper.session.IMAGENET_MEAN: tuple[float, float, float] = (0.485, 0.456, 0.406)

Per-channel mean the 1.x models were trained against.

Torchvision’s standard ImageNet normalisation, applied after a bicubic resize to 64x64. The v2 BlinkCNN uses plain /255 instead; the two are not interchangeable, and applying the wrong one degrades accuracy without erroring.

blinklinmult.paper.session.IMAGENET_STD: tuple[float, float, float] = (0.229, 0.224, 0.225)

Per-channel standard deviation. See IMAGENET_MEAN.

blinklinmult.paper.session.CROPS_INPUT = 'crops'

Name of the eye-crop input in every exported graph.

blinklinmult.paper.session.FEATURES_INPUT = 'features'

Name of the descriptor input, present only in the two-stream graph.

blinklinmult.paper.session.SEQUENCE_OUTPUT = 'sequence'

Name of the per-frame output in every exported graph.

blinklinmult.paper.session.CLIP_OUTPUT = 'clip'

Name of the pooled output, present only in the two-stream graph.

exception blinklinmult.paper.session.SessionError

Bases: Exception

Raised when a graph cannot be loaded or is fed the wrong inputs.

Initialize self. See help(type(self)) for accurate signature.

class blinklinmult.paper.session.OnnxModel(session: onnxruntime.InferenceSession, metadata: dict[str, Any] | None = None)

A frozen 1.x model, backed by an onnxruntime session.

Parameters:

• session (ort.InferenceSession) – The loaded graph.

• metadata (dict[str, Any]) – Its sidecar, if one was found beside it.

property needs_features: bool

Whether this graph takes the 160-d descriptor stream.

__call__(crops: numpy.ndarray, features: numpy.ndarray | None = None) → dict[str, numpy.ndarray]

Run the graph.

Parameters:

• crops (np.ndarray) – (B, T, 3, H, W) float32, already normalised.

• features (np.ndarray | None) – (B, T, 160) float32, for the two-stream graph.

Returns:

Output name to array. Always carries

"sequence" (B, T, 1); the two-stream graph also carries "clip" (B, 1).

Return type:

dict[str, np.ndarray]

Raises:

SessionError – If a required input is missing.

blinklinmult.paper.session.load_onnx(path: str | pathlib.Path, providers: list[str] | None = None) → OnnxModel

Open an exported graph.

Parameters:

• path (str | Path) – The .onnx file. A .onnx.json sidecar beside it is read if present.

• providers (list[str] | None) – Execution providers, in preference order. Defaults to onnxruntime’s own choice.

Returns:

Ready to run.

Return type:

OnnxModel

Raises:

SessionError – If the file is missing, or onnxruntime is not installed.