Tip

This tutorial is available as a Jupyter notebook.

🤗 HuggingFace Models

This tutorial will demonstrate how to fine-tune a pretrained HuggingFace transformer using the composer library! Composer provides a highly optimized training loop and the ability to compose several methods that can accelerate training.

We will focus on fine-tuning a pretrained BERT-base model on the Stanford Sentiment Treebank v2 (SST-2) dataset. After fine-tuning, the BERT model should be able to determine if a setence has positive or negative sentiment.

Let’s do this 🚀

Install Composer

To use HuggingFace with Composer, we’ll need to install Composer with the NLP dependencies. If you haven’t already, run:

%pip install 'mosaicml[nlp]'

Import Hugging Face pretrained model

First, we import a pretrainec BERT model and tokenizer from the transformers library. We alter the model to output two classes for sentiment classification by setting num_labels=2.

import transformers

# Create a BERT sequence classification model using HuggingFace transformers
model = transformers.AutoModelForSequenceClassification.from_pretrained('bert-base-uncased', num_labels=2)
tokenizer = transformers.AutoTokenizer.from_pretrained('bert-base-uncased')

Creating dataloaders

Next, we will download and tokenize the SST-2 datasets.

import datasets
from multiprocessing import cpu_count

# Create BERT tokenizer
def tokenize_function(sample):
    return tokenizer(
        text=sample['sentence'],
        padding="max_length",
        max_length=256,
        truncation=True
    )

# Tokenize SST-2
sst2_dataset = datasets.load_dataset("glue", "sst2")
tokenized_sst2_dataset = sst2_dataset.map(tokenize_function,
                                          batched=True,
                                          num_proc=cpu_count(),
                                          batch_size=100,
                                          remove_columns=['idx', 'sentence'])

# Split dataset into train and validation sets
train_dataset = tokenized_sst2_dataset["train"]
eval_dataset = tokenized_sst2_dataset["validation"]

Here, we will create a PyTorch DataLoader for each of the datasets generated in the previous block.

from torch.utils.data import DataLoader
data_collator = transformers.data.data_collator.default_data_collator
train_dataloader = DataLoader(train_dataset, batch_size=16, shuffle=False, drop_last=False, collate_fn=data_collator)
eval_dataloader = DataLoader(eval_dataset,batch_size=16, shuffle=False, drop_last=False, collate_fn=data_collator)

Use HuggingFaceModel to convert model to ComposerModel

from torchmetrics import Accuracy
from torchmetrics.collections import MetricCollection
from composer.models.huggingface import HuggingFaceModel
from composer.metrics import CrossEntropy

metrics = [CrossEntropy(), Accuracy()]
# Package as a composer model
composer_model = HuggingFaceModel(model, metrics=metrics, use_logits=True)

Optimizers and Learning Rate Schedulers

The last setup step is to create an optimizer and a learning rate scheduler. We will use PyTorch’s AdamW optimizer and linear learning rate scheduler since these are typically used to fine-tune BERT on tasks such as SST-2.

from torch.optim import AdamW
from torch.optim.lr_scheduler import LinearLR

optimizer = AdamW(
    params=composer_model.parameters(),
    lr=3e-5, betas=(0.9, 0.98),
    eps=1e-6, weight_decay=3e-6
)
linear_lr_decay = LinearLR(
    optimizer, start_factor=1.0,
    end_factor=0, total_iters=150
)

Composer Trainer

We will now specify a Composer Trainer object and run our training! Trainer has many arguments that are described in our documentation, so we’ll discuss only the less-obvious arguments used below: - max_duration - a string specifying how long to train, either in terms of batches (e.g. ‘10ba’ is 10 batches) or epochs (e.g. ‘1ep’ is 1 epoch). - schedulers - a list of PyTorch learning rate schedulers that will be composed together. - device - specifies if the training will be done on CPU or GPU by using ‘cpu’ or ‘gpu’, respectively. - train_subset_num_batches - specifies the number of training batches to use for each epoch. This is not a necessary argument but is useful for quickly testing code. - precision - whether to do the training in full precision ('fp32') or mixed precision ('amp'). Mixed precision provides an almost 2x speedup in training time on certain hardware. - seed - sets the random seed for the training run, so the results are reproducible!

import torch
from composer import Trainer

# Create Trainer Object
trainer = Trainer(
    model=composer_model,
    train_dataloader=train_dataloader,
    eval_dataloader=eval_dataloader,
    max_duration="1ep",
    optimizers=optimizer,
    schedulers=[linear_lr_decay],
    device='gpu' if torch.cuda.is_available() else 'cpu',
    train_subset_num_batches=150,
    precision='fp32',
    seed=17
)
# Start training
trainer.fit()

Visualizing Results

To check the training’s validation accuracy, we read the Trainer object state.current_metrics['eval']

trainer.state.current_metrics['eval']

Our model reaches ~86% accuracy with only 150 iterations of training! Let’s visualize a few samples from the validation set to see how our model performs.

eval_batch = next(iter(eval_dataloader))

# Move batch to gpu
eval_batch = {k: v.cuda() if torch.cuda.is_available() else v for k, v in eval_batch.items()}
with torch.no_grad():
    predictions = composer_model(eval_batch)["logits"].argmax(dim=1)

# Visualize only 5 samples
predictions = predictions[:6]

label = ['negative', 'positive']
for i, prediction in enumerate(predictions[:6]):
    sentence = sst2_dataset["validation"][i]["sentence"]
    correct_label = label[sst2_dataset["validation"][i]["label"]]
    prediction_label = label[prediction]
    print(f"Sample: {sentence}")
    print(f"Label: {correct_label}")
    print(f"Prediction: {prediction_label}")
    print()

Save Fine-Tuned Model

Finally, to save the fine-tuned model parameters we call PyTorch save method and pass it the model’s state_dict:

torch.save(trainer.state.model.state_dict(), 'model.pt')

Conclusion

This tutorial showed how to use the Composer Trainer to fine-tune a pre-trained BERT on a subset of the SST-2 dataset. We focused on Composer’s basic functionality, but there are many more features, such as easy-to-use gradient accumulation and multi-GPU training! Check out our documentation for more info.