Beginner question: Can I use W&B without modifying my training script much?

Great question - it's honestly one of W&B's best features. You really can just sprinkle in a few lines. The main bits are importing wandb, calling wandb.init() at the start, and then wandb.log() inside your training loop to send metrics over. It takes maybe five minutes.

It doesn't feel intrusive at all. Your core training logic stays completely intact, you're just adding lightweight calls to ship data out. Give it a shot with a simple script - you'll be surprised how quick it is to get rolling.

Totally! That's exactly why it caught on so quickly in my team. You keep your training loop's core logic untouched, just decorate it with a few wandb calls. It's almost like adding print statements but way more powerful.

For a super basic PyTorch example, imagine your loop looks like this now:
```python
for epoch in range(epochs):
train_loss = train_one_epoch(...)
val_loss = validate(...)
print(f"Epoch {epoch}: train_loss {train_loss}, val_loss {val_loss}")
```

Adding W&B literally just wraps those print statements:
```python
import wandb
wandb.init(project="my_project")

for epoch in range(epochs):
train_loss = train_one_epoch(...)
val_loss = validate(...)
wandb.log({"train_loss": train_loss, "val_loss": val_loss})
```

The only "intrusive" part I've found is if you later want to log gradients or hyperparameters, but even that's just a config object passed to `init`. Start with just `log` and you're golden.

The minimal approach described works well for logging basic metrics. Where I find teams eventually make more changes is when they start tracking hardware utilization or dataset versions alongside those metrics. For that, you might add a couple more config parameters to wandb.init.

But to start? Just those three lines. The real value is that it scales from there without forcing a rewrite. You can add artifact logging for your model checkpoints later with maybe two more lines, keeping the same init and log structure.