How to Use wandb for Tracking Machine Learning Experiments

To use wandb for tracking, first initialize a project with wandb.init(). Then log metrics or parameters during training using wandb.log(). This helps you monitor and visualize your model's progress easily.

Syntax

Here is the basic syntax to use wandb for tracking your machine learning experiments:

• wandb.init(project='project_name'): Starts a new run in the specified project.
• wandb.config: Stores hyperparameters or settings.
• wandb.log({'metric_name': value}): Logs metrics like loss or accuracy during training.
• wandb.finish(): Ends the run and uploads data.

python

import wandb

# Start a new run
wandb.init(project='my_project')

# Set hyperparameters
wandb.config.learning_rate = 0.01

# Log metrics
wandb.log({'loss': 0.5, 'accuracy': 0.8})

# Finish the run
wandb.finish()

Example

This example shows how to track a simple training loop with wandb. It logs loss and accuracy for 5 steps.

python

import wandb
import random

# Initialize wandb run
wandb.init(project='example_project')

# Set hyperparameters
wandb.config.batch_size = 32
wandb.config.epochs = 1

# Simulate training loop
for step in range(5):
    loss = random.uniform(0.1, 1.0)  # fake loss
    accuracy = random.uniform(0.5, 1.0)  # fake accuracy
    wandb.log({'loss': loss, 'accuracy': accuracy, 'step': step})

# Finish run
wandb.finish()

Output

wandb: Currently logged in as: <your_username>
wandb: Tracking run with wandb version 0.15.4
wandb: Run data is saved locally and will be synced to the cloud.
wandb: Run URL: https://wandb.ai/<your_username>/example_project/runs/<run_id>

Common Pitfalls

Common mistakes when using wandb include:

• Not calling wandb.init() before logging, which causes errors.
• Forgetting to call wandb.finish(), so data may not upload properly.
• Logging metrics without unique step values, making plots confusing.
• Not setting a project name, which scatters runs across default projects.

Always initialize wandb at the start, log with step info, and finish the run.

python

import wandb

# Wrong: logging before init
# wandb.log({'loss': 0.5})  # This will error

# Right way:
wandb.init(project='fix_pitfalls')
wandb.log({'loss': 0.5, 'step': 1})
wandb.finish()

Quick Reference

Command	Purpose
wandb.init(project='name')	Start a new tracking run in a project
wandb.config	Store hyperparameters or settings
wandb.log({'metric': value})	Log metrics or values during training
wandb.finish()	End the run and upload data

Key Takeaways

Always start with wandb.init() to create a new tracking run.

Use wandb.log() to record metrics like loss and accuracy during training.

Set wandb.config to save hyperparameters for easy comparison.

Call wandb.finish() to properly end and upload your run data.

Include step numbers in logs to keep metric plots clear and ordered.