WeightsLab: An Open-Source Tool for Visualizing Neural Network Training Processes

WeightsLab is an open-source experimental management tool for deep learning researchers. It supports real-time monitoring of model parameters, data samples, and training metrics, and provides interactive editing capabilities to help developers more intuitively understand and optimize the neural network training process. This tool addresses the core challenge of understanding and debugging training processes in deep learning.

Existing deep learning tools like Weights & Biases or TensorBoard often lack fine-grained insight into the model's internal state during training. WeightsLab, developed by the Graybx team and released in late 2024 under the Apache 2.0 license, was created to solve this problem. Its core positioning is to be a 'visual debugger for neural networks', allowing researchers to understand the model's learning process like stepping through code in an IDE.

WeightsLab integrates with training code via a lightweight Python SDK without changing the original logic. Its core capabilities include: 1. Fine-grained data insight: Track individual samples to see their contribution to loss and identify anomalies. 2. Real-time parameter monitoring: Capture and display changes in weights, biases, and optimizer states to detect issues like gradient vanishing. 3. Interactive experiment editing: Modify hyperparameters, remove low-quality samples, or adjust model architecture during training to shorten iteration cycles.

To get started with WeightsLab: 1. Ensure Python version is 3.10-3.15, then install via pip install weightslab. 2. Launch the visual interface with weightslab ui launch. 3. Integrate into training code by importing weightslab and wrapping model, optimizer, loss function, and data loader with the watch_or_edit method (non-intrusive, no code refactoring needed).

Compared to tools like Weights & Biases, WeightsLab's unique value lies in real-time interaction (vs traditional black-box recorders). It supports migration from wandb to reduce switching costs. Its local-first design stores data locally by default with optional cloud sync, suitable for sensitive data or offline work.

WeightsLab includes an AI assistant for natural language interaction (data sorting, filtering, diagnosis) with two modes: local Ollama deployment or cloud OpenRouter (default Llama3.3 70B). The data slicing feature allows creating custom subsets based on sample features, loss contribution, or confidence, useful for active learning and data cleaning.

WeightsLab prioritizes security: default non-secure mode, with TLS encryption and gRPC authentication support. Generate certificates via weightslab se command, stored in .weightslab-certs folder, meeting both local development and production security needs.

Graybx is building a community via Slack and GitHub Discussions. Future plans include more powerful model architecture editing (pruning, growth). Existing examples cover image classification, segmentation, and detection with PyTorch/PyTorch Lightning. WeightsLab shifts from passive recording to active interaction, making deep learning development more transparent, controllable, and efficient.