Adafactor with Accelerator

You've demonstrated a proof-of-concept of Adafactor with Trainer to train your language translation model with reduced memory requirements. Now you'd like to customize your training loop using Accelerator. Build the training loop to use Adafactor!

The compute_optimizer_size() function has been pre-defined. Some training objects have been pre-loaded: model, train_dataloader, and accelerator. Adafactor has been pre-imported from torch.optim.

本练习是课程的一部分

Efficient AI Model Training with PyTorch

查看课程

练习说明

Pass the model parameters to Adafactor when defining the optimizer.
Pass in the optimizer state to print the size.

交互式实操练习

通过完成这段示例代码来试试这个练习。

# Pass the model parameters to Adafactor
optimizer = Adafactor(params=____.____())
model, optimizer, train_dataloader = accelerator.prepare(model, optimizer, train_dataloader)

for batch in train_dataloader:
    inputs, targets = batch["input_ids"], batch["labels"]
    outputs = model(inputs, labels=targets)
    loss = outputs.loss
    accelerator.backward(loss)
    optimizer.step()
    optimizer.zero_grad()

# Pass in the optimizer state
total_size_megabytes, total_num_elements = compute_optimizer_size(optimizer.state.____())
print(f"Number of optimizer parameters: {total_num_elements:,}\nOptimizer size: {total_size_megabytes:.0f} MB")

编辑并运行代码