Skip to main content

Hook-Based Feature Extractor

You do not need to rewrite a model to see inside it.

Key Insight​

A forward hook is a callback you register on any nn.Module. PyTorch calls it automatically after that module's forward pass, passing in the input and output tensors. You can capture the output — called activations — without touching the model's code at all.

Why This Matters​

Feature extraction and visualization are essential for understanding what a network has learned. Hooks let you tap into any layer of any pretrained model in just a few lines, making them the standard tool for interpretability and transfer learning.