experiment-tracking-swanlab

To track ML experiment metrics in a self-hosted dashboard, initialize a run with swanlab.init and log metrics like train/loss and val/accuracy using swanlab.log. This centralizes your experiment context for reliable comparison across local and self-hosted environments.

Yes, you can use SwanLab for experiment tracking with PyTorch Lightning by integrating the SwanLabLogger. This framework-native integration minimizes manual boilerplate by automatically capturing training and evaluation metrics in your dashboard.

The best way to log media assets like images and audio during model training is using swanlab.log. It attaches images, audio, text notes, GIF/video, and point clouds to the same run, requiring the pillow and soundfile dependencies for better debugging and reporting.

SwanLab supports offline experiment tracking via mode="local". This allows offline-first metrics logging without immediate cloud connectivity, enabling later inspection and synchronization using swanlab watch and swanlab sync commands.

Yes, you need extra dependencies to log audio and images to SwanLab runs. Installing the core swanlab package must be accompanied by pillow for image processing and soundfile for audio processing to enable lightweight media logging.

To integrate experiment tracking into a Hugging Face Transformers Trainer workflow, set report_to="swanlab" in your training arguments. This automatically logs training and evaluation metrics to your SwanLab run dashboard without manual logging boilerplate.