Weights & Biases (wandb)

Warnings

• breaking Python 3.8 is no longer supported starting from `wandb` version 0.25.0.
  Fix: Upgrade your Python environment to 3.9 or higher.
• breaking The legacy `wandb.beta.workflows` module (including `log_model()`, `use_model()`, `link_model()`) was removed in version 0.24.0. These functions are no longer available and will cause `AttributeError`.
  Fix: Migrate to the modern artifact API using `Run.log_artifact()`, `Run.use_artifact()`, and `Run.link_artifact()` methods.
• breaking Version `0.24.0` was yanked from PyPI due to a critical bug that could cause silent failure to upload some run data. If used, data might be missing from your W&B dashboard.
  Fix: Immediately upgrade to `wandb` version `0.24.1` or higher. Missing data from `0.24.0` runs can often be recovered by running `wandb sync` on the `.wandb` files.
• deprecated Several `wandb.Run` methods are deprecated in favor of direct properties, including `run.project_name()`, `run.get_url()`, `run.get_project_url()`, and `run.get_sweep_url()`.
  Fix: Use the direct properties instead: `run.project`, `run.url`, `run.project_url`, and `run.sweep_url` respectively.
• gotcha The `wandb: ERROR Run aborted` or `wandb: ERROR Failed to log data` messages indicate an unexpected termination or data logging failure. This can be caused by script errors, system resource constraints, or network connectivity issues.
  Fix: Check script for exceptions, monitor system resources (CPU/RAM), verify stable network connection, and ensure data logged to `wandb.log()` is in the correct dictionary format.
• gotcha Programmatic dataset splitting (e.g., using `sklearn.model_selection.train_test_split` without fixing `random_state` or without managing splits as artifacts) can lead to inconsistent train/test sets when new data is added, invalidating comparisons between experiments.
  Fix: Ensure reproducibility of splits (e.g., set `random_state` or explicitly manage dataset versions as W&B Artifacts) to maintain consistent evaluation benchmarks across experiments.

Install

• pip install wandb Install latest version

Imports

• wandb

  import wandb

• wandb.init

  import wandb
  wandb.init(...)

• wandb.login

  import wandb
  wandb.login()

Quickstart

This quickstart demonstrates how to initialize a Weights & Biases run, log hyperparameters using `wandb.config`, and track metrics like loss and accuracy using `run.log()` within a simulated training loop. Before running, ensure you have authenticated with `wandb.login()` or set the `WANDB_API_KEY` environment variable.

import wandb
import os

# Authenticate with W&B. For automated environments, use an environment variable.
# wandb.login() will prompt for an API key if not set.
# Set WANDB_API_KEY environment variable for CI/CD or headless environments.
# For local development, running `wandb login` in your terminal is common.
# os.environ.get('WANDB_API_KEY', '') # Example for fetching from env, but wandb.login() handles this.
wandb.login()

# Initialize a new W&B run
project_name = os.environ.get('WANDB_PROJECT', 'my-awesome-project')
config = {
    'epochs': 10,
    'lr': 0.01,
    'batch_size': 32
}

with wandb.init(project=project_name, config=config) as run:
    # Access hyperparameters
    epochs = run.config.epochs
    learning_rate = run.config.lr

    print(f"Starting training for {epochs} epochs with LR: {learning_rate}")

    # Simulate a training loop
    for epoch in range(epochs):
        # Simulate loss and accuracy metrics
        loss = 1.0 / (epoch + 1) + 0.1 * (epochs - epoch - 1) / epochs
        accuracy = 0.5 + 0.5 * (epoch + 1) / epochs

        # Log metrics to W&B
        run.log({"epoch": epoch, "loss": loss, "accuracy": accuracy})
        print(f"Epoch {epoch+1}/{epochs}: Loss = {loss:.4f}, Accuracy = {accuracy:.4f}")

    print("Training complete!")