Track Your First ML Experiment with Weights & Biases

1. 1
   Install & Login

   Install the Python library and authenticate your machine with your Weights & Biases account. The login command will prompt you for an API key, which you can find in your W&B profile settings.
2. 2
   Initialize a Run

   In your training script, import `wandb` and call `wandb.init()` to start a new experiment run. This creates a unique, trackable run within a specified project on your Weights & Biases dashboard.
3. 3
   Track Hyperparameters

   Pass a dictionary of hyperparameters to `wandb.init()` via the `config` argument. This saves the configuration with the run, making it easy to compare different experimental setups later.
4. 4
   Log Metrics

   Inside your model's training loop, call `wandb.log()` with a dictionary of metrics like loss and accuracy. Weights & Biases automatically plots these values over time (by default, against the step number).
5. 5
   View Your Dashboard

   After your script runs, follow the link printed in your terminal or navigate to your project page on the Weights & Biases website. You'll find interactive charts, hyperparameter tables, and system metrics for your run.

import wandb
import random
import time

# 1. Start a new run
wandb.init(
    project="my-first-project",
    # Track hyperparameters and run metadata
    config={
        "learning_rate": 0.02,
        "architecture": "dummy-NN",
        "dataset": "synthetic",
        "epochs": 5,
    }
)

# This simple loop simulates a training process
epochs = wandb.config.epochs
learning_rate = wandb.config.learning_rate
offset = random.random() / 5

print(f"Starting training for {epochs} epochs...")

for epoch in range(1, epochs + 1):
    # Simulate training step
    time.sleep(1)
    
    # 2. Log metrics to W&B
    acc = 1 - (1 / (epoch + offset)) - random.random() / epoch - learning_rate
    loss = (1 / (epoch + offset)) + random.random() / epoch + learning_rate
    
    wandb.log({"accuracy": acc, "loss": loss, "epoch": epoch})
    print(f"Epoch {epoch}: Accuracy={acc:.4f}, Loss={loss:.4f}")

# 3. Mark the run as finished
wandb.finish()

print("Run complete. Check your Weights & Biases dashboard!")