Repository files navigation

🧠 Simple utility functions and wrappers for hooking onto layers within PyTorch models for feature extraction.

pip install torch-featurelayer

Imports:

importtorchfromtorchvision.modelsimportvgg11fromtorch_featurelayerimportFeatureLayer

Load a pretrained VGG-11 model:

model=vgg11(weights='DEFAULT').eval()

Hook onto layerfeatures.15 of the model:

layer_path='features.15'hooked_model=FeatureLayer(model,layer_path)

Forward pass an input tensor through the model:

x=torch.randn(1,3,224,224)feature_output,output=hooked_model(x)

feature_output is the output of layerfeatures.15. Print the output shape:

print(f'Feature layer output shape:{feature_output.shape}')# [1, 512, 14, 14]print(f'Model output shape:{output.shape}')# [1, 1000]

Check theexamples directory for more.

TheFeatureLayer class wraps a model and provides a hook to access the output of a specific feature layer.

• __init__(self, model: torch.nn.Module, feature_layer_path: str)

  Initializes theFeatureLayer instance.

  • model: The model containing the feature layer.
  • feature_layer_path: The path to the feature layer in the model.

• __call__(self, *args: Any, **kwargs: Any) -> tuple[torch.Tensor | None, torch.Tensor]

  Performs a forward pass through the model and updates the hooked feature layer.

  • *args: Variable length argument list.
  • **kwargs: Arbitrary keyword arguments.

  Returns a tuple containing the feature layer output and the model output.

TheFeatureLayers class wraps a model and provides hooks to access the output of multiple feature layers.

• __init__(self, model: torch.nn.Module, feature_layer_paths: list[str])

  Initializes theFeatureLayers instance.

  • model: The model containing the feature layers.
  • feature_layer_paths: A list of paths to the feature layers in the model.

• __call__(self, *args: Any, **kwargs: Any) -> tuple[dict[str, torch.Tensor | None], torch.Tensor]

  Performs a forward pass through the model and updates the hooked feature layers.

  • *args: Variable length argument list.
  • **kwargs: Arbitrary keyword arguments.

  Returns a tuple containing the feature layer outputs and the model output.

Theget_layer_candidates function returns a generator of layer paths for a given model up to a specified depth.

• model: The model to get layer paths from.
• max_depth: The maximum depth to traverse in the model's layers.

Returns a generator of layer paths.

MIT