Nowadays, it's extremely common to find yourself working with large language models daily, either as a tool to aid you in your job (e.g., ChatGPT) or because you work with the models themselves, as part of an application you are developing (e.g., a chatbot), or, like myself, you are doing research on natural language processing or other similar domains.
If you are currently working or researching with large language models, there are two ways to go about it: you either develop/research on top of some API provided by OpenAI, Google, Microsoft, etc.; or you have to work on the models by yourself (e.g. training, fine-tuning, doing inference on top of them, etc.). This article might be right for those facing the latter scenario. I will present two essential tools I use in my day-to-day job that give me all the flexibility I require to do my job right and how you can combine these tools to your advantage.
Hugging Face
When you have to train or fine-tune large language models by yourself, as I do, a great tool you might have encountered at some point is Hugging Face [1], particularly their Transformers API [2]. This company provides access to thousands of state-of-the-art pre-trained models [3] and provides APIs to make it easy for anyone to train, share, and build anything on top of them.
Hugging Face provides the interfaces for various models across different tasks: natural language processing, computer vision, audio, and multimodal. It supports different back-end tools such as PyTorch [4], JAX [5], and TensorFlow [6]. It provides a large community of developers and extensive documentation [7] over its APIs.
PyTorch Lightning
PyTorch Lightning [8] started as a library on top of PyTorch that provided the necessary framework to abstract and simplify the process of training models. They have evolved and eventually became Lightning AI [9], a company dedicated to simplifying the process of training, evaluating, deploying, and maintaining deep learning models. What I like most about this framework is the seamless integration it provides in order to use multiple GPUs [10], experiment tracking tools [11], and the overall simplification of the whole training and evaluation process.
Why should we use Hugging Face with PyTorch Lightning?
If we're being honest, the API provided by Hugging Face [12] for fine-tuning their models is usually a good enough solution for most cases. In particular, if you are not trying to fine-tune any model and use the inference, their pipeline API [13] is more than enough. Moreover, their Trainer API and library should be more than enough when dealing with some classical tasks such as text classification. And it has been constantly upgrading, so it's highly likely that in the future this will be improved even further.
But there is a reality: the library has to be compatible with (at least so far) 3 different deep learning frameworks: PyTorch, TensorFlow, and JAX. Consequently, you won't be able to fully utilize certain features from PyTorch if you limit yourself to training and fine-tuning the models only using their API.
In my case, I find their version of Multiple GPUs training and parallelism a little confusing (but maybe that's me because I have been working with Lightning for quite some time now). On the other hand, I like the Lighting integration [14] with MLFlow [15], another tool I use on a day-to-day basis because it makes the job of tracking experiments much easier.
Training Hugging Face Transformers with PyTorch Lightning
Installation
Before we delve into how to train a Hugging Face transformer using the framework provided by Pytorch Lightning, we need to install the Hugging Face and the Lightning libraries:
pip install lightning transformers datasetsThis will install the two libraries we need for development and, for this specific case, the library datasets [16] from Hugging Face as well, which offers a wide variety of datasets for you to experiment with and share.
LightningDataModule
We start by defining the LightningDataModule [17]. This is an abstraction provided by PyTorch Lightning that encapsulates all the steps needed to process a dataset: download, tokenize, clean, transform and any other form of pre-processing is done within this module:
from datasets import load_dataset
from lightning import LightningDataModule
from transformers import AutoTokenizer
class HFDataModule(LightningDataModule):
def __init__(self, tokenizer_name):
super().__init__()
self.tokenizer_name = tokenizer_name
def _tokenize_function(self, batch):
return self.tokenizer(batch["text"],
padding="max_length",
truncation=True,
max_length=64)
def prepare_data(self):
# Download the dataset and tokenizer
load_dataset("rotten_tomatoes")
self.tokenizer = AutoTokenizer.from_pretrained(self.tokenizer_name)
def setup(self, stage):
dataset = load_dataset("rotten_tomatoes")
dataset = dataset.rename_column("label", "labels")
self.dataset = dataset.map(self._tokenize_function, batched=True)
self.dataset.set_format(
type="torch",
columns=["input_ids", "token_type_ids", "attention_mask", "labels"]
)
def train_dataloader(self):
return DataLoader(self.dataset["train"],
batch_size=32,
shuffle=True)
def val_dataloader(self):
return DataLoader(self.dataset["validation"],
batch_size=32,
shuffle=False)
def test_dataloader(self):
return DataLoader(self.dataset["test"],
batch_size=32,
shuffle=False)The HFDataModule class takes the name of the Hugging Face tokenizer (which is usually the same one as the model) and loads the "rotten_tomatoes" dataset [18] from Hugging Face. In the __init__ method, the most important part is calling super() in order to set up the base LightningDataModule class.
The prepare_data() is a method utilized by the Lightning Trainer (which we'll see in the next sections) and is useful for things as data download: in this case, the method calls the load_dataset() method from Hugging Face, but only to download the dataset's files into the system, it also downloads the tokenizer's files and loads it.
The setup() is another method called by the Trainer, usually for pre-processing and splitting the data. It takes a stage argument that can be 'fit', 'validate', 'test', or 'predict', depending on what the Trainer is going to do. In this case, we ignore it and set the Hugging Face dataset (which is already split into train, test, and validation).
First, the dataset is loaded into an attribute, renaming the column "label" to "labels" (this is because Hugging Face expects the name to be "labels" but when using their APIs that rename is handled internally), it tokenizes the text, sets the format of the tensors to PyTorch and selects the columns that the Hugging Face model needs.
Finally, each <split>_dataloader method loads the DataLoaders for each training split.
LightningModule
Next, we build the LightningModule [19], which is going to be the model to train. In this case, we set a module that is basically a wrapper around the Hugging Face Model for Sentence Classification [20]. The advantage here is defining a subclass of LightningModule, which is a subclass of PyTorch's nn.Module, gives us absolute freedom regarding what to use with the model: extra layers, optimization, scheduler, loss function, etc.
from lightning import LightningModule
from torch.optim import AdamW
from transformers import AutoModelForSequenceClassification
class TransformerModule(LightningModule):
def __init__(self, model_name):
super().__init__()
self.save_hyperparameters()
self.model = AutoModelForSequenceClassification.from_pretrained(
model_name
)
def forward(self, **inputs):
return self.model(**inputs)
def training_step(self, batch, batch_idx):
loss = self(**batch).loss
self.log("train_loss", loss)
return loss
def validation_step(self, batch, batch_idx):
loss = self(**batch).loss
self.log("val_loss", loss, on_epoch=True)
return loss
def test_step(self, batch, batch_idx):
loss = self(**batch).loss
self.log("test_loss", loss, on_epoch=True)
return loss
def configure_optimizers(self):
return AdamW(self.parameters(), lr=1e-3)With the aid of Lightning, we have a fully functional model ready for training. It loads the Hugging Face model and declares the corresponding steps of train, test, and validation, logging the loss in each of them. We also set up the optimizer to use.
Training and Evaluation
Now that we have the Data Module and the Model itself, we need to set up the training loop and the evaluation. But that's where all of PyTorch Lightning's magic happens:
from lightning import Trainer
from lightning.pytorch.callbacks import EarlyStopping, ModelCheckpoint
MODEL_NAME = "bert-base-uncased"
dm = HFDataModule(MODEL_NAME)
model = TransformerModule(MODEL_NAME)
early_stopping = EarlyStopping(monitor="val_loss", patience=3)
model_checkpoints = ModelCheckpoint("./checkpoints",
monitor="val_loss",
save_top_k=1)
trainer = Trainer(max_epochs=5, callbacks=[early_stopping, model_checkpoints])
trainer.fit(model, datamodule=dm)
trainer.test(ckpt_path=model_checkpoints.best_model_path, datamodule=dm)The previous snippet runs the whole training process using the "bert-base-uncased" model from Hugging Face, and fine-tuning it for the dataset of "rotten_tomatoes", which is a sentiment classification dataset over movie reviews.
First, we set up the data module and the models using "bert-base-uncased" as our base model, which loads the correct model and tokenizer.
We continue by creating an instance of the Trainer [21]. This is the class behind all of PyTorch Lightning's magic, it is the one that will take the LightningModule and run the training, validation, and test loops over the different splits given by the LightningDataModule. In this case, we limit the initialization parameters to the number of epochs and we add a couple of callbacks: The early stopping callback monitors the metric given by the parameter, this metric should be logged by the LightningModule's log() method (in this case it will look for the "val_loss" that is logged in the corresponding validation_step that is part of the TransformerModule class we defined above); the ModelCheckpoint callback is used to store the best checkpoint result based on the monitored metric (i.e., "val_loss").
The Trainer module is a very versatile and powerful abstraction of the training loop, it provides all the necessary tools to run multiple GPUs (or other types of accelerators), setup different loggers, setting up a maximum amount of time, steps or epochs for running the training loop, etc.
The fit() method takes the model and data module we created before and runs for a maximum of 5 epochs (or will stop before if the early stopping condition is met).
We could use the save_checkpoint method of the Trainer to save a checkpoint after the training stops. In this case, however, since we have the ModelCheckpoint callback, we know that the best model is saved in model_checkpoints.best_model_path.
We use the test method on the best model checkpoint to run the evaluation over the test data.
Although Lightning has access to many different loggers, by default, it will run a TensorBoardLogger [22] under the "lightning_logs" directory (in the same directory where the script was run), and by default, Lightning AI will install TensorBoard. Thus, we can check the logged results:
tensorboard --logdir ./lightning_logsWe now have the possibility of running multiple experiments and easily comparing the results of these experiments with the help of TensorBoard.
Final Thoughts
Hugging Face has an excellent community and repository of pre-trained models of a very different nature and with a lot of different possibilities to explore. They also offer a suite of datasets already available to download, most of them pre-processed and already split in train, test, and validation. They offer a nice set of APIs for running pipelines and fine-tuning any of the models they provide. However, in my experience, their API is a little tricky and difficult to use, especially when having to go beyond what is doable out of the box. Besides, as there are multiple supported backends, it is harder to take advantage of all the possibilities a specific backend offers, in our case, PyTorch.
PyTorch Lightning provides a very good wrapper around PyTorch, giving their users access to lots of abstractions that simplify the training, testing, and validation process. Their LightningModule has access to all of PyTorch technology, which basically renders full flexibility in using any loss, optimizer, scheduler, etc., from PyTorch. Their Trainer module grants users access to different things to maximize the deep learning workflow, such as multi-GPU training, early stopping, or model checkpointing.
If we combine them, we can have all the power of Hugging Face's Transformers with all the flexibility and scalability of the PyTorch Lightning framework—a real winning combination.
References
[1] Hugging Face Inc. https://huggingface.co/
[2] Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumond, Clement Delangue, Anthony Moi, Pierric Cistac, Tim Rault, Remi Louf, Morgan Funtowicz, Joe Davison, Sam Shleifer, Patrick von Platen, Clara Ma, Yacine Jernite, Julien Plu, Canwen Xu, Teven Le Scao, Sylvain Gugger, et al.. 2020. Transformers: State-of-the-Art Natural Language Processing. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, pages 38–45, Online. Association for Computational Linguistics.
[3] Hugging Face Inc. Models. https://huggingface.co/models
[4] PyTorch. https://pytorch.org/
[5] JAX: High-Performance Array Computing. https://jax.readthedocs.io/en/latest/
[6] TensorFlow. https://www.tensorflow.org/
[7] Hugging Face Inc. Documentation. https://huggingface.co/docs
[8] Lightning AI. PyTorch Lightning Documentation. https://lightning.ai/docs/pytorch/stable/
[9] Lightning AI. https://lightning.ai/
[10] Lightning AI. GPU Training (Expert). https://lightning.ai/docs/pytorch/stable/accelerators/gpu_expert.html
[11] Lightning AI. Track and Visualize Experiments (Advanced). https://lightning.ai/docs/pytorch/stable/visualize/logging_advanced.html#logger
[12] Hugging Face Inc. Fine-tune a pretrained model. https://huggingface.co/docs/transformers/training
[13] Hugging Face Inc. Pipelines for inference. https://huggingface.co/docs/transformers/pipeline_tutorial
[14] Lightning AI. MLFlowLogger Documentation. https://lightning.ai/docs/pytorch/stable/extensions/generated/lightning.pytorch.loggers.MLFlowLogger.html
[15] MLFlow. ML and GenAI made simple. https://mlflow.org/
[16] Hugging Face Inc. Datasets. https://huggingface.co/datasets
[17] Lightning AI. LightningDataModule. https://lightning.ai/docs/pytorch/stable/data/datamodule.html
[18] Hugging Face Inc. Datasets: rotten_tomatoes. https://huggingface.co/datasets/rotten_tomatoes
[19] Lightning AI. LightningModule. https://lightning.ai/docs/pytorch/stable/common/lightning_module.html
[20] Hugging Face Inc. AutoModelForSequenceClassification. https://huggingface.co/docs/transformers/v4.37.2/en/model_doc/auto#transformers.AutoModelForSequenceClassification
[21] Lightning AI. Trainer. https://lightning.ai/docs/pytorch/stable/common/trainer.html
[22] Lightning AI. TensorBoardLogger. https://lightning.ai/docs/pytorch/stable/api/lightning.pytorch.loggers.tensorboard.html
ความคิดเห็น