weights-and-biases

Call wandb.init with your project name and config dict, then call wandb.log inside your training loop with metrics like loss and accuracy. Finish with wandb.finish, and use wandb.save or Artifacts to upload model checkpoints.

Define a sweep_config with a search method (grid, random, or bayes), a target metric, and parameter distributions, then create it with wandb.sweep. Launch trials with wandb.agent pointing to a train function that reads hyperparameters from wandb.config.

Yes. Set report_to="wandb" in TrainingArguments for automatic HuggingFace logging, or pass WandbLogger to the PyTorch Lightning Trainer. Keras uses dedicated callbacks from wandb.integration.keras such as WandbMetricsLogger.

Artifacts version any file-based asset like datasets or checkpoints with automatic lineage tracking between runs. The Model Registry is a curated layer where model artifacts are linked and promoted through aliases like staging and production.

Yes. Set WANDB_MODE=offline before initializing, and all metrics are written locally. Sync them to the cloud later by running wandb sync on the run directory once connectivity is restored.

Use grid search for a few discrete parameters needing full coverage, random search for quick exploration of many parameters, and Bayesian optimization when training runs are expensive and you need sample-efficient search. Bayesian is the recommended default.