Show HN: WeightsLab – turn training signals into a map of your bad data

Category: devtools

Tags: dataset-debugging, pytorch, computer-vision

Score: 7.0/10 (Innovation: 6, Technical: 7, Documentation: 8, Utility: 7)

WeightsLab is an open-source PyTorch tool that connects live training loss signals back to individual data samples, enabling real-time detection of mislabels, outliers, and class imbalance in computer vision datasets. Its ability to pause, curate, and resume training without restarting offers a novel debugging workflow that goes beyond standard experiment logging tools like Weights & Biases.

Target audience: machine learning engineers, data scientists, computer vision researchers

Repository: https://github.com/GrayboxTech/weightslab · Python · Apache-2.0 · 126 stars

View on Hacker News