Weights & Biases로 실험 추적하기

You will run a model dozens of times with tiny tweaks. Without records you forget what worked. Experiment tracking logs every run so progress is never lost. 📊

Weights and Biases, often written wandb, is a popular tool that records metrics, settings, and outputs from each training run in one dashboard.

One call opens a new tracked session. wandb.init names your project so every run lands in the same place for easy comparison.

Save the settings that shaped a run, like learning rate and batch size. Storing this config lets you tie any result back to its exact recipe.

Inside your loop, send numbers you care about. wandb.log records loss and accuracy so they plot as live curves over time.

Every logged number streams to a web dashboard. You watch training curves update in real time and spot a diverging run before it wastes hours.

The real power is comparison. Overlay many runs and the dashboard shows which hyperparameters actually moved your accuracy upward.

Beyond numbers you can store files. Logging a trained checkpoint as an artifact keeps the exact weights paired with the run that made them.

When training ends, close the session so all data flushes to the server. Calling wandb.finish marks the run complete and ready to review.

Tired of tuning by hand? A wandb sweep launches many runs across a range of settings and reports which combination wins automatically.

Logged runs become shared history. Teammates open the same dashboard and instantly see your results, making your work reproducible and easy to trust.

You learned to track runs with wandb: init a project, log config and metrics, save artifacts, and finish. Now every experiment is recorded. 🎉