VisualTorch

[]() []()

VisualTorch aims to help visualize Torch-based neural network architectures. It currently supports generating flow-style, graph-style, and LeNet-style architectures for PyTorch Sequential and Custom models. Its original visual styles were inspired by visualkeras, pytorchviz, pytorch-summary, and torchview; since then, it has grown its own unified tracing backend and architecture-handling logic well beyond its origins.

Note: 1.0+ is a major release with breaking API changes, but with significantly better features and algorithms - upgrading is recommended. For the old API, use 0.2.5 or older.

Limitation: VisualTorch traces a real forward pass to build the diagram, which has an inherent limitation shared by any tracing-based approach (not a bug, and not fixable without full symbolic execution): models with data-dependent control flow (e.g. a branch only taken if a tensor value crosses some threshold) only show whichever branch the traced dummy input happened to take. Separately, a layer that returns multiple meaningful output tensors (e.g. a custom multi-task head, or nn.LSTM's (output, (h_n, c_n))) still has its node's size based on only its first tensor; with show_dimension=True, every output tensor's shape is shown in the label, not just the first. Downstream connections are correct either way. Contributions are welcome!

Animated Reveal

Every style can also render as an animated GIF, revealing the model one layer/column at a time, via visualtorch.animate(model, input_shape, style=...) - see it in action for flow, graph, and lenet styles.

Documentation

Online documentation is available at visualtorch.readthedocs.io.

The docs include usage examples, API references, and other useful information.

Installation

See the Installation page.

MCP integration

VisualTorch includes an optional, client-neutral stdio MCP server for generating static PNG diagrams and animated GIF reveals from PyTorch model source. It exposes capability discovery, structured output metadata, documentation resources, subprocess timeouts, and all three canonical styles (graph, flow, and lenet). Install it with pip install "visualtorch[mcp]" and see the MCP integration guide for the tool schemas, generic stdio configuration, examples, and trusted-code security boundary. No client-specific plugin or extension is required.

Used in Research

VisualTorch has been used in published research, including works published in Nature, IEEE, and MDPI.

See the Research Showcase page for the full list.

Used VisualTorch in your research, built something with it, or found a paper that cites it? Tell us about it or open a pull request to add it directly - we'd love to hear.

Examples

See the Usage Examples page.

Contributing

Please feel free to send a pull request to contribute to this project by following this guideline.

Releases

See GOVERNANCE.md for release methodology and cadence, and the PyPI release history for past releases.

License

This poject is available as open source under the terms of the MIT License.

Originally, this project was based on the visualkeras (under the MIT license), with additional modifications inspired by pytorchviz, pytorch-summary, and torchview, all of which are also licensed under the MIT license.

Citation

Please cite this project in your publications if it helps your research.

Note: the paper below describes VisualTorch as of its publication date (2024). The project has since been substantially refactored, including breaking API changes (see the documentation for the current API) - the DOI always resolves to what was actually reviewed and published.

@article{Hendria2024,
  doi = {10.21105/joss.06678},
  url = {https://doi.org/10.21105/joss.06678},
  year = {2024},
  publisher = {The Open Journal},
  volume = {9},
  number = {102},
  pages = {6678},
  author = {Willy Fitra Hendria and Paul Gavrikov},
  title = {VisualTorch: Streamlining Visualization for PyTorch Neural Network Architectures},
  journal = {Journal of Open Source Software}
}

Star History