• feature_extraction
  • ResnetEmbeddingsExtractor
    • __init__
    • to_device
    • forward
    • from_dataloader
  • concatenate_layers
  • concatenate_two_layers

Provides classes and functions for extracting embedding vectors from neural networks.

class ResnetEmbeddingsExtractor(torch.nn.Module)

A class to hold, and extract embedding vectors from, a resnet.

Attributes:

• backbone - The resnet from which to extract embedding vectors.

 | __init__(backbone_name: str, device: torch.device) -> None

Construct the backbone and set appropriate mode and device

Arguments:

• backbone_name - The name of the desired backbone. Must be one of: [resnet18, wide_resnet50].
• device - The device where to run the network.

 | to_device(device: torch.device) -> None

Perform device conversion on backone

See pytorch docs for documentation on torch.Tensor.to

 | forward(batch: torch.Tensor, channel_indices: Optional[torch.Tensor] = None, layer_hook: Optional[Callable[[torch.Tensor], torch.Tensor]] = None, layer_indices: Optional[List[int]] = None) -> torch.Tensor

Run inference on backbone and return the embedding vectors.

Arguments:

• batch - A batch of images.
• channel_indices - A list of indices with the desired channels to include in the embedding vectors.
• layer_hook - A function that runs on each layer of the resnet before concatenating them.
• layer_indices - A list of indices with the desired layers to include in the embedding vectors.

Returns:

• embedding_vectors - The embedding vectors.

 | from_dataloader(dataloader: DataLoader, channel_indices: Optional[torch.Tensor] = None, layer_hook: Optional[Callable[[torch.Tensor], torch.Tensor]] = None, layer_indices: Optional[List[int]] = None) -> torch.Tensor

Same as self.forward but take a dataloader instead of a tensor as argument.

concatenate_layers(layers: List[torch.Tensor]) -> torch.Tensor

Scale all tensors to the heigth and width of the first tensor and concatenate them.

concatenate_two_layers(layer1: torch.Tensor, layer2: torch.Tensor) -> torch.Tensor

Scale the second tensor to the height and width of the first tensor and concatenate them.