FeatLens

FeatLens offers a powerful, model-agnostic feature map visualization tool specifically designed for vision models. With capabilities to render feature maps from various well-known architectures such as DINO, DINOv2, DINOv3, CLIP, and more, FeatLens provides insights into the inner workings of deep learning models by allowing users to analyze features from any layer and any model source including timm, Hugging Face transformers, and external repositories.

Key Features

• Comprehensive Model Support: Visualize feature maps from a wide range of vision models including DINO, CLIP, MAE, and more.
• Flexible Layer Access: Easily extract and view feature maps from any layer of the selected model, providing a customizable experience.
• Diverse Visualization Methods: Leverage various robust visualization techniques including Principal Component Analysis (PCA), cosine similarity, k-means segmentation, and saliency mapping. This allows detailed representation of features and their importance across selected layers.
• Cross-Model Comparison: Compare the output of different models side by side effortlessly, facilitating meaningful insights into model performance and architecture differences.

Example Usage

To visualize feature maps from a model, the following code snippet illustrates how to use FeatLens effectively:

import featlens as ll

# Visualize feature maps using the DINO model at specified layers.
ll.visualize("dinov2_vitb14", "img.jpg", layers=[2, 5, 8, 11], out="row.png")

Visual Comparison of Models

An example demonstrates how to compare models at their final layer:

# Compare multiple models at the final layer.
ll.compare(["dino_vitb16", "mae_vitb16", "clip_large_openai"], "img.jpg", layer=-1, out="cmp.png")

Gallery of Visualizations

The gallery section shows a variety of visual outputs including feature maps for images like peacocks, cats, and market scenes, showcasing how different models respond to the same input image:

Advanced Features

• Attention Rollout: Analyze where the model focuses its attention within an image.
• Semantic Correspondence: Find matches of specific features across different images, which can highlight similarities in patterns or objects.
• Video Feature Extraction: Process video clips frame by frame to visualize model features over time.

By utilizing FeatLens, researchers and developers can better understand the nuances of vision models, leading to improved model development and deployment strategies. For full documentation and additional examples, please visit the official documentation.

Explore the feature maps of any vision model in-depth with FeatLens, enhancing the interpretability of deep learning algorithms.