willyfh/visualtorch

What it solves

VisualTorch provides a way to visualize PyTorch neural network architectures, making it easier to understand and document the structure of both Sequential and Custom models. It addresses the problem of complex model architectures that are difficult to visualize manually or through simple text logs.

How it works

The tool uses a tracing-based approach, performing a real forward pass with a dummy input to map out the model's connections and layers. It can generate diagrams in three distinct styles: flow-style, graph-style, and LeNet-style. Additionally, it supports creating animated GIF reveals that show the model architecture layer-by-layer.

Who it’s for

Researchers and developers working with PyTorch who need to clear, professional diagrams of their neural network architectures for research papers, documentation, or debugging.

Highlights

• Multiple Visualization Styles: Supports flow, graph, and LeNet-style diagrams.
• Animated Reveals: Ability to render the model architecture as an animated GIF.
• ** uma Unified Tracing Backend**: Features its own tracing logic to handle PyTorch models.
• MCP Integration: Includes an optional MCP server for generating diagrams from model source via stdio.
• Research Proven: Used in publications across Nature, IEEE, and MDPI.