Comgra is a Python tool for PyTorch users to analyze and debug neural networks by visualizing computation graphs and tensor data through a GUI. It aids both novice users in understanding network internals and experienced professionals in advanced analysis and anomaly detection, offering a comprehensive view of network behavior.

How It Works

Comgra records tensor data and network structure during training. Its core is a dependency graph visualizing tensor relationships, allowing users to navigate the computation flow. Selectors enable granular inspection of tensors, displaying summary statistics or individual neuron values. This approach simplifies complex graphs by abstracting away unnecessary details, making it easier to compare architectures and trace issues.

Quick Start & Requirements

• Install via pip: pip install comgra
• Requires Python.
• Run GUI: comgra --path "/path/to/your/data/group_name"
• Example usage and Colab notebook available in the README.

Highlighted Details

• Interactive GUI with dependency graph, tensor selectors, and data output.
• Supports custom visualizations via a --visualization-file parameter.
• Dynamic recording allows retroactive selection of batches and samples for analysis.
• Detailed tutorial demonstrates finding a subtle bug related to activation functions.

Maintenance & Community

• Developed by Florian Dietz.
• Contact: floriandietz44@gmail.com for future development contributions.

Licensing & Compatibility

• License not explicitly stated in the README.

Limitations & Caveats

• Requires requires_grad=True on registered tensors, which may have side effects if directly manipulated.
• Integer tensors or tensors where requires_grad cannot be set require workarounds using add_tensor_connection.
• Dynamic recordings can introduce a performance overhead.