dataeval.extractors.TorchExtractor¶

class dataeval.extractors.TorchExtractor(model, transforms=None, device=None, layer_name=None, use_output=True)¶

Extracts embeddings from a PyTorch model, with optional intermediate layer hooking.

Encapsulates all PyTorch-specific logic for feature extraction:

• Model management (torch.nn.Module)

• Device handling

• Transform pipeline

• Layer hooking for intermediate layer extraction

Implements the FeatureExtractor protocol.

Parameters:¶

model : torch.nn.Module¶

PyTorch model for feature extraction.

transforms : Transform or Sequence[Transform] or None, default None¶

Preprocessing transforms to apply before encoding. When None, uses raw images.

device : DeviceLike or None, default None¶

Device for computation. When None, uses DataEval’s configured device.

layer_name : str or None, default None¶

Layer to extract embeddings from. When None, uses model output.

use_output : bool, default True¶

If True, captures layer output; if False, captures layer input. Only used when layer_name is specified.

Example

Basic usage with a model:

>>> import torch.nn as nn
>>> from dataeval import Embeddings
>>> from dataeval.extractors import TorchExtractor
>>>
>>> model = nn.Sequential(nn.Flatten(), nn.Linear(784, 128))
>>> extractor = TorchExtractor(model, device="cpu")
>>> embeddings = Embeddings(dataset, extractor=extractor, batch_size=32)

Extracting from an intermediate layer:

>>> extractor = TorchExtractor(
...     model,
...     layer_name="0",  # Extract from Flatten layer
...     use_output=True,
... )

property layer_name : str | None¶

Return the layer name for intermediate extraction, if set.

property use_output : bool¶

Return whether output (True) or input (False) is captured from the layer.