Training loops before and after Accelerator

You want to modify a PyTorch training loop to use Accelerator for your language model to simplify translations using the MPRC dataset of sentence paraphrases. Update the training loop to prepare your model for distributed training.

Some data has been pre-loaded:

accelerator is an instance of Accelerator
train_dataloader, optimizer, model, and lr_scheduler have been defined and prepared with Accelerator

Deze oefening maakt deel uit van de cursus

Efficient AI Model Training with PyTorch

Cursus bekijken

Oefeninstructies

Update the .to(device) lines so that Accelerator handles device placement.
Modify the gradient computation to use Accelerator.

Praktische interactieve oefening

Probeer deze oefening eens door deze voorbeeldcode in te vullen.

for batch in train_dataloader:
    optimizer.zero_grad()
    inputs, targets = batch["input_ids"], batch["labels"]
    # Update the lines so Accelerator handles device placement
    inputs = inputs.to(device)
    targets = targets.to(device)
    outputs = model(inputs, labels=targets)
    loss = outputs.loss
    # Modify the gradient computation to use Accelerator
    ____.backward(____)
    optimizer.step()
    lr_scheduler.step()

Code bewerken en uitvoeren