algaradi/resvisualize

Easy and Useful ResNet Visualization Tool for AI Researchers and Developers

ResNet50 Feature Map Visualization Tool

A comprehensive visualization tool for exploring and understanding how ResNet50 processes images through its internal layers.

Features

Multiple Visualization Techniques

• Feature Maps: Visualize activation patterns at different layers
• Activation Maximization: Generate images that maximally activate specific neurons
• Layer-wise Relevance Propagation (LRP): Understand which input pixels contribute most to specific activations
• Grad-CAM: Visualize important regions for classification decisions

Advanced Customization Options

• Image Preprocessing:

  • Resize with multiple strategies (simple, preserve aspect ratio, pad and resize)
  • Custom cropping options (center, random, five-crop, ten-crop)
  • Normalization options (ImageNet standard, simple, custom)
  • Data augmentation (color jitter, rotation, flips, grayscale, blur)

• Visualization Controls:

  • Custom grid size for feature map displays
  • Channel selection across all network layers
  • Multiple normalization methods for visualization
  • Various colormaps for different visualization preferences

Analysis Tools

• Channel statistics and distribution visualizations
• Feature map overlays on original images
• Export options (JSON, CSV, image downloads)
• Detailed processing logs

Installation

1. Clone the repository:

git clone https://github.com/algaradi/resvisualize.git
cd ResVisualize

2. Install the required Python dependencies:

pip install -r requirements.txt

Usage

1. Start the Flask server:

python app.py

2. Open your browser and navigate to:

http://localhost:5000

3. Upload an image and experiment with different visualization techniques!

Visualization Techniques Explained

Feature Maps

Shows the actual activations of different channels in each layer when processing an input image. This helps understand what patterns each filter responds to.

Activation Maximization

Generates synthetic images that maximally activate specific neurons, revealing what patterns or features a particular filter is looking for.

Layer-wise Relevance Propagation (LRP)

Highlights which parts of the input image contribute most to specific activations or classifications, helping explain model decisions.

Grad-CAM

Creates class activation maps showing which regions of an image are important for classification decisions.

System Requirements

• Python 3.7+
• PyTorch 2.0+
• Modern web browser with JavaScript enabled
• GPU recommended for faster processing (but not required)

Technical Details

• Backend: Flask (Python)
• Frontend: HTML, JavaScript, TailwindCSS
• Deep Learning: PyTorch, Captum
• Visualization: Matplotlib, Seaborn