dreamlens 0.1.1

PyTorch tools for activation atlases, feature visualization, and image amplification.

DreamLens

Feature visualization in native PyTorch. DreamLens works with torchvision and other compatible torch.nn.Module models through one FeatureVisualizer API.

Install

pip install dreamlens

Install the notebook dependencies with:

pip install "dreamlens[examples]"

API

from torchvision.models import ResNet18_Weights, resnet18
from dreamlens import FeatureTarget, FeatureVisualizer

model = resnet18(weights=ResNet18_Weights.DEFAULT).eval()
visualizer = FeatureVisualizer(model, normalize=True)
target = FeatureTarget.for_channel("layer2.1.conv2", 17, reduction="norm")

result = visualizer.visualize(target, method="maximize")
result.save("channel_17.png")

The same class provides four workflows:

visualizer.visualize(target, method="maximize")
visualizer.visualize(target, method="maco")
visualizer.visualize(
    target,
    method="feature_accentuation",
    image="image.jpg",
    regularization_layer="layer2.1",
)
visualizer.visualize(
    method="caricature",
    image="image.jpg",
    layers=["layer3.1.conv2"],
)

The notebooks contain the complete configurations for each method.

Results

Activation maximization

MaCo

Feature accentuation

Caricature

Notebooks

• Getting started
• Activation maximization
• MaCo
• Feature accentuation
• Caricature
• Result gallery

Research

DreamLens brings ideas from three papers into one PyTorch API:

• Feature Visualization, Distill 2017
• MAgnitude Constrained Optimization, NeurIPS 2023
• Feature Accentuation, 2024

This is an independent project for education and research. It is not an official implementation from the paper authors.

License

DreamLens code is Apache-2.0 licensed. Papers, model weights, datasets and example images keep their original rights and licenses. Educational use does not remove those obligations. See NOTICE for sources and attribution.