Neural Network Trainer
This project is a web-based neural network training tool built with Blazor, designed to help users create, train, and test fully connected neural networks (excluding CNNs and RNNs). The backend uses TensorFlow.NET to handle model construction and training logic, while the front end provides an intuitive interface for configuring network parameters and monitoring results in real time.
Users can upload CSV datasets, define features and labels, configure model layers and activation functions, and specify training settings such as epochs and batch size. During training, the system displays real-time metrics (loss, accuracy, validation curves) using MudBlazor charts. Once training is complete, users can test the model in an interactive playground by entering custom input values to view predictions.
Key features & challenges
- Upload CSV datasets and define features and labels
- Configure model layers, activation functions, and training settings (epochs, batch size)
- Real-time training metrics (loss, accuracy, validation curves) via MudBlazor charts
- Interactive playground to test trained models with custom input values
My contributions
- Set up the project structure and integrated MudBlazor for a responsive UI
- Built backend RESTful API endpoints for file upload, dataset retrieval, and training initiation
- Implemented Entity Framework for persistent dataset storage, including dataset history management
- Developed major front-end pages: main navigation, dataset configuration, neural network parameter, and uploaded dataset management pages
- Configured the GitHub repository and handled project organization