Gradient accumulation with Accelerator

You're training a language model to simplify translations by paraphrasing complex sentences, but your GPU is running out of memory. Gradient accumulation allows the model to effectively train on larger batches by using small batches that fit into memory. You prefer to write the training loop explicitly to see its structure, so you're using Accelerator. Note that this exercise actually runs on the CPU, but the code remains the same for the GPU.

The model, train_dataloader, optimizer, and lr_scheduler have been pre-defined.

This exercise is part of the course

Efficient AI Model Training with PyTorch

View Course

Exercise instructions

Configure Accelerator() to use gradient accumulation with two steps.
Set up an Accelerator context manager to enable gradient accumulation for the model.

Hands-on interactive exercise

Have a go at this exercise by completing this sample code.

# Configure Accelerator
accelerator = ____(____=____)
model, optimizer, train_dataloader, lr_scheduler = accelerator.prepare(model, optimizer, train_dataloader, lr_scheduler)
for batch in train_dataloader:
    # Set up an Accelerator context manager
    with ____.____(____):
        inputs, targets = batch["input_ids"], batch["labels"]
        outputs = model(inputs, labels=targets)
        loss = outputs.loss
        accelerator.backward(loss)
        optimizer.step()
        lr_scheduler.step()
        optimizer.zero_grad()
        print(f"Loss = {loss}")

Edit and Run Code