experiment-tracking-swanlab

What problem does it solve? Machine learning training runs produce scattered metrics, configs, and artifacts that are hard to compare and reproduce. This Skill provides guidance for logging experiments with SwanLab, an open-source tracker that works in cloud, local, or self-hosted modes. ## Core Features & Use Cases - Run and Config Tracking: Initialize runs with swanlab.init, capture hyperparameters, and log scalar metrics with stable namespaced keys like train/loss and val/accuracy. - Media and Chart Logging: Log images, audio, text, GIF videos, point clouds, molecules, and custom swanlab.echarts visualizations such as line, bar, and heatmap charts. - Framework Integrations: Connect with PyTorch, HuggingFace Transformers via report_to="swanlab" or SwanLabCallback, PyTorch Lightning via SwanLabLogger, and Fastai via SwanLabCallback. - Use Case: A researcher fine-tuning a BERT classifier can set report_to="swanlab" in TrainingArguments, then compare validation loss across seeds in the SwanLab dashboard, or run fully offline with mode="local" and inspect logs later using swanlab watch. ## Quick Start Use the SwanLab skill to set up experiment tracking for my PyTorch training loop with config capture, loss logging, and a local dashboard.