TransViz - Neural Network Visualization Toolkit

STILLS UNDER EARLY PHASES OF DEVELOPMENT

Work being done:

• Packaging and dependencies
• Tests & Examples of Usage
• Issues being found at usage
• Overall the python engine

Work not yet done:

• UI stills in just first draft
• CI/CD, etc

A real-time visualization and debugging toolkit for neural networks, featuring:

• 🎮 Interactive tensor inspection
• ⏯️ Training process time travel
• 🔍 Breakpoint debugging system
• 📊 Real-time metrics dashboard
• 🔗 Distributed training support

Key Features

• Real-time Visualization: Monitor activations/gradients during forward/backward passes
• Breakpoint System: Set conditional breakpoints with tensor pattern matching
• Metrics Dashboard: Track loss, accuracy, and custom metrics in real-time
• Differential Updates: Efficient binary diffs for large tensor visualization
• Multi-Framework Support: Designed for PyTorch with extensibility for other frameworks

Installation

# Clone repository
git clone https://github.com/fblgit/transviz.git
cd transviz

# Install dependencies
pip install -r requirements.txt

Quick Start

from transviz import visualize, VizConfig

# Instrument your model
@visualize.trace(name="transformer_block")
class TransformerBlock(nn.Module):
    def forward(self, x):
        x = self.attention(x)
        visualize.breakpoint("post_attention", x)
        return self.ffn(x)

# Initialize visualizer
viz = visualize.ModelVisualizer(
    VizConfig(
        port=8080,
        mode="hybrid",
        capture_gradients=True
    )
)
viz.start()

# Train your model normally
train(model, dataloader)

Access the visualization dashboard at http://localhost:8080

Configuration

@dataclass
class VizConfig:
    port: int = 8080              # Web server port
    mode: str = "hybrid"          # light|hybrid|debug
    storage_path: str = "./viz_data"
    tensor_sampling: float = 0.1  # Fraction of data to capture
    capture_gradients: bool = False
    breakpoint_timeout: int = 300  # seconds

API Reference

Decorators

@visualize.trace(name="layer", capture_shapes=True)
@visualize.breakpoint("attention_weights", condition=lambda t: t.max() > 1.0)

Core Methods

visualize.log_metrics({"loss": 0.5})
visualize.set_breakpoint("gradient_check", condition=has_nan)
visualize.save_session("training_run_1")

CLI Commands

transviz serve --port 8080 --model ./checkpoint.pth
transviz export --session latest --format hdf5

Contributing

1. Fork the repository
2. Create feature branch (git checkout -b feature/amazing-feature)
3. Commit changes (git commit -m 'Add amazing feature')
4. Push to branch (git push origin feature/amazing-feature)
5. Open Pull Request

Roadmap
[ ] TensorFlow/Keras support
[ ] Jupyter notebook integration
[ ] Pre-trained model zoo integration
[ ] 3D tensor visualization