🌊 NeuralWave

Real-time Neural Network Model Analysis & 3D Visualization

NeuralWave is a web-based tool for benchmarking, analyzing, and visualizing neural network models. It provides deep insights into model architecture through interactive 3D visualizations, allowing you to inspect individual layers, neurons, and connections.

---

✨ Features

📊 Model Management

• Auto-Discovery: Scans configured directories for .gguf model files
• Verification: Tests if models can be loaded successfully
• Benchmarking: Measures tokens-per-second (TPS) performance
• Batch Operations: Verify, benchmark, or analyze multiple models at once

🔬 Weight Analysis

• Layer Statistics: Mean, StdDev, Min, Max for each layer's weights
• Parameter Counts: Total params per layer and model-wide
• Architecture Detection: Identifies MLP, Attention, Norm, Embedding layers

🎮 3D Visualization

• Interactive Scene: Orbit, pan, zoom through the neural network
• Layer Selection: Click layers or use dropdown to focus
• Deep Dive: Expand any layer to see its internal structure:
  • Neurons with value-based coloring
  • Weight connections
  • Biases visualization

🔬 Feature Extraction

• Attention Layers: Visualize all 12 heads with Q/K/V neurons
• MLP Layers: See Gate, Up, Down projections separately
• Norm Layers: Scale (γ) and Shift (β) parameters
• One-Click: Press "Extract Features" button to decompose

---

🚀 Quick Start

Prerequisites

• Go 1.21+
• LOOM - OpenFluke's neural network framework
• Hugging Face account (for downloading models)

Installation

git clone https://github.com/openfluke/neuralwave.git
cd neuralwave
go mod tidy
go build -o neuralwave .

Download Models

Use the included script to download models from Hugging Face:

./download_models.sh

Or manually download SafeTensors models to your ~/.cache/huggingface/hub/ directory.

Run

./neuralwave

Open http://localhost:3000 in your browser.

---

📁 Project Structure

neuralwave/
├── main.go              # CLI entry point & llama.cpp integration
├── web.go               # HTTP server, WebSocket, API handlers
├── views/
│   └── index.html       # Single-page app with Three.js visualization
├── static/              # Static assets
├── download_models.sh   # Model downloader script
├── go.mod
└── README.md

---

🎯 Usage

Dashboard Tab

• View all discovered models
• Select models for batch operations
• Run Verify / Benchmark / Analyze

Analysis Tab

• See detailed layer-by-layer statistics
• Compare model architectures

Compare Tab

• Side-by-side layer comparison across models
• Filter by layer name or type

Visualize 3D Tab

1. Select a model from the dropdown
2. Click Generate to render the architecture
3. Click any layer to zoom in and see details
4. Press 🔬 Extract Features for deep decomposition
5. Rotate/Zoom freely to explore
6. Click ✕ to return to overview

---

⚙️ Configuration

Edit main.go to configure:

modelRoots := []string{
    "/path/to/your/models",
}

---

🛠️ Tech Stack

• Backend: Go, Gorilla WebSocket
• AI Framework: LOOM (OpenFluke)
• Frontend: HTML, CSS (Tailwind), JavaScript
• 3D Engine: Three.js (r128)
• Model Format: SafeTensors (Hugging Face)

---

📝 License

APACHE2 License - See LICENSE for details.

---

🤝 Contributing

Contributions welcome! Please open an issue or PR.

---

Built with 🧠 by OpenFluke