SmartNetToy

Visual Neural Architecture Search

SmartNetToy is a visual, interactive neural architecture search (NAS) tool that combines a drag-and-drop network builder with automated search algorithms guided by SHAP (SHapley Additive exPlanations) and Fisher Information metrics. The tool democratizes neural architecture design by making it accessible to researchers and engineers who may not have deep expertise in architecture engineering.

The drag-and-drop interface allows users to visually compose neural network architectures by connecting layer blocks (convolution, attention, pooling, normalization, etc.) on a canvas. The system automatically validates architecture constraints (tensor shape compatibility, memory budget, latency targets) in real-time as blocks are connected, preventing invalid configurations before training.

The automated NAS mode uses evolutionary search with SHAP-guided mutation operators that preferentially modify layers with high attribution scores and Fisher Information-guided crossover that preserves the most informative layer combinations. This informed search converges 3-5x faster than random mutation baselines to architectures of equivalent quality.

Built in approximately 4,800 lines of Python with a Flask web interface and PyTorch backend, SmartNetToy supports one-click training, real-time loss visualization, and export to standard PyTorch module code for production deployment.

• Drag-and-drop neural architecture builder with real-time constraint validation
• SHAP + Fisher Information guided evolutionary NAS, 3-5x faster convergence
• ~4,800 LOC with Flask web UI and PyTorch training backend
• One-click training with real-time loss visualization
• Export to standard PyTorch module code for production deployment