Fine-tuning a pretrained model

In this tutorial, we will show you how to fine-tune a pretrained model from the Transformers library. In TensorFlow, models can be directly trained using Keras and the fit method. In PyTorch, there is no generic training loop so the 🤗 Transformers library provides an API with the class Trainer to let you fine-tune or train a model from scratch easily. Then we will show you how to alternatively write the whole training loop in PyTorch.

Before we can fine-tune a model, we need a dataset. In this tutorial, we will show you how to fine-tune BERT on the IMDB dataset: the task is to classify whether movie reviews are positive or negative. For examples of other tasks, refer to the additional-resources section!

Preparing the datasets

We will use the 🤗 Datasets library to download and preprocess the IMDB datasets. We will go over this part pretty quickly. Since the focus of this tutorial is on training, you should refer to the 🤗 Datasets documentation or the preprocessing tutorial for more information.

First, we can use the load_dataset function to download and cache the dataset:

from datasets import load_dataset

raw_datasets = load_dataset("imdb")

This works like the from_pretrained method we saw for the models and tokenizers (except the cache directory is ~/.cache/huggingface/dataset by default).

The raw_datasets object is a dictionary with three keys: "train", "test" and "unsupervised" (which correspond to the three splits of that dataset). We will use the "train" split for training and the "test" split for validation.

To preprocess our data, we will need a tokenizer:

from transformers import AutoTokenizer

tokenizer = AutoTokenizer.from_pretrained("bert-base-cased")

As we saw in preprocessing, we can prepare the text inputs for the model with the following command (this is an example, not a command you can execute):

inputs = tokenizer(sentences, padding="max_length", truncation=True)

This will make all the samples have the maximum length the model can accept (here 512), either by padding or truncating them.

However, we can instead apply these preprocessing steps to all the splits of our dataset at once by using the map method:

def tokenize_function(examples):
    return tokenizer(examples["text"], padding="max_length", truncation=True)

tokenized_datasets = raw_datasets.map(tokenize_function, batched=True)

You can learn more about the map method or the other ways to preprocess the data in the 🤗 Datasets documentation.

Next we will generate a small subset of the training and validation set, to enable faster training:

small_train_dataset = tokenized_datasets["train"].shuffle(seed=42).select(range(1000)) 
small_eval_dataset = tokenized_datasets["test"].shuffle(seed=42).select(range(1000)) 
full_train_dataset = tokenized_datasets["train"]
full_eval_dataset = tokenized_datasets["test"]

In all the examples below, we will always use small_train_dataset and small_eval_dataset. Just replace them by their full equivalent to train or evaluate on the full dataset.

Fine-tuning in PyTorch with the Trainer API

Since PyTorch does not provide a training loop, the 🤗 Transformers library provides a Trainer API that is optimized for 🤗 Transformers models, with a wide range of training options and with built-in features like logging, gradient accumulation, and mixed precision.

First, let’s define our model:

from transformers import AutoModelForSequenceClassification

model = AutoModelForSequenceClassification.from_pretrained("bert-base-cased", num_labels=2)

This will issue a warning about some of the pretrained weights not being used and some weights being randomly initialized. That’s because we are throwing away the pretraining head of the BERT model to replace it with a classification head which is randomly initialized. We will fine-tune this model on our task, transferring the knowledge of the pretrained model to it (which is why doing this is called transfer learning).

Then, to define our Trainer, we will need to instantiate a TrainingArguments. This class contains all the hyperparameters we can tune for the Trainer or the flags to activate the different training options it supports. Let’s begin by using all the defaults, the only thing we then have to provide is a directory in which the checkpoints will be saved:

from transformers import TrainingArguments

training_args = TrainingArguments("test_trainer")

Then we can instantiate a Trainer like this:

from transformers import Trainer

trainer = Trainer(
    model=model, args=training_args, train_dataset=small_train_dataset, eval_dataset=small_eval_dataset
)

To fine-tune our model, we just need to call

trainer.train()

which will start a training that you can follow with a progress bar, which should take a couple of minutes to complete (as long as you have access to a GPU). It won’t actually tell you anything useful about how well (or badly) your model is performing however as by default, there is no evaluation during training, and we didn’t tell the Trainer to compute any metrics. Let’s have a look on how to do that now!

To have the Trainer compute and report metrics, we need to give it a compute_metrics function that takes predictions and labels (grouped in a namedtuple called EvalPrediction) and return a dictionary with string items (the metric names) and float values (the metric values).

The 🤗 Datasets library provides an easy way to get the common metrics used in NLP with the load_metric function. here we simply use accuracy. Then we define the compute_metrics function that just convert logits to predictions (remember that all 🤗 Transformers models return the logits) and feed them to compute method of this metric.

import numpy as np
from datasets import load_metric

metric = load_metric("accuracy")

def compute_metrics(eval_pred):
    logits, labels = eval_pred
    predictions = np.argmax(logits, axis=-1)
    return metric.compute(predictions=predictions, references=labels)

The compute function needs to receive a tuple (with logits and labels) and has to return a dictionary with string keys (the name of the metric) and float values. It will be called at the end of each evaluation phase on the whole arrays of predictions/labels.

To check if this works on practice, let’s create a new Trainer with our fine-tuned model:

trainer = Trainer(
    model=model,
    args=training_args,
    train_dataset=small_train_dataset,
    eval_dataset=small_eval_dataset,
    compute_metrics=compute_metrics,
)
trainer.evaluate()

which showed an accuracy of 87.5% in our case.

If you want to fine-tune your model and regularly report the evaluation metrics (for instance at the end of each epoch), here is how you should define your training arguments:

from transformers import TrainingArguments

training_args = TrainingArguments("test_trainer", evaluation_strategy="epoch")

See the documentation of TrainingArguments for more options.

Fine-tuning with Keras

Models can also be trained natively in TensorFlow using the Keras API. First, let’s define our model:

import tensorflow as tf
from transformers import TFAutoModelForSequenceClassification

model = TFAutoModelForSequenceClassification.from_pretrained("bert-base-cased", num_labels=2)

Then we will need to convert our datasets from before in standard tf.data.Dataset. Since we have fixed shapes, it can easily be done like this. First we remove the “text” column from our datasets and set them in TensorFlow format:

tf_train_dataset = small_train_dataset.remove_columns(["text"]).with_format("tensorflow")
tf_eval_dataset = small_eval_dataset.remove_columns(["text"]).with_format("tensorflow")

Then we convert everything in big tensors and use the tf.data.Dataset.from_tensor_slices method:

train_features = {x: tf_train_dataset[x] for x in tokenizer.model_input_names}
train_tf_dataset = tf.data.Dataset.from_tensor_slices((train_features, tf_train_dataset["label"]))
train_tf_dataset = train_tf_dataset.shuffle(len(tf_train_dataset)).batch(8)

eval_features = {x: tf_eval_dataset[x] for x in tokenizer.model_input_names}
eval_tf_dataset = tf.data.Dataset.from_tensor_slices((eval_features, tf_eval_dataset["label"]))
eval_tf_dataset = eval_tf_dataset.batch(8)

With this done, the model can then be compiled and trained as any Keras model:

model.compile(
    optimizer=tf.keras.optimizers.Adam(learning_rate=5e-5),
    loss=tf.keras.losses.SparseCategoricalCrossentropy(from_logits=True),
    metrics=tf.metrics.SparseCategoricalAccuracy(),
)

model.fit(train_tf_dataset, validation_data=eval_tf_dataset, epochs=3)

With the tight interoperability between TensorFlow and PyTorch models, you can even save the model and then reload it as a PyTorch model (or vice-versa):

from transformers import AutoModelForSequenceClassification

model.save_pretrained("my_imdb_model")
pytorch_model = AutoModelForSequenceClassification.from_pretrained("my_imdb_model", from_tf=True)

Fine-tuning in native PyTorch

You might need to restart your notebook at this stage to free some memory, or execute the following code:

del model
del pytorch_model
del trainer
torch.cuda.empty_cache()

Let’s now see how to achieve the same results as in trainer section in PyTorch. First we need to define the dataloaders, which we will use to iterate over batches. We just need to apply a bit of post-processing to our tokenized_datasets before doing that to:

• remove the columns corresponding to values the model does not expect (here the "text" column)
• rename the column "label" to "labels" (because the model expect the argument to be named labels)
• set the format of the datasets so they return PyTorch Tensors instead of lists.

Our tokenized_datasets has one method for each of those steps:

tokenized_datasets = tokenized_datasets.remove_columns(["text"])
tokenized_datasets = tokenized_datasets.rename_column("label", "labels")
tokenized_datasets.set_format("torch")

small_train_dataset = tokenized_datasets["train"].shuffle(seed=42).select(range(1000))
small_eval_dataset = tokenized_datasets["test"].shuffle(seed=42).select(range(1000))

Now that this is done, we can easily define our dataloaders:

from torch.utils.data import DataLoader

train_dataloader = DataLoader(small_train_dataset, shuffle=True, batch_size=8)
eval_dataloader = DataLoader(small_eval_dataset, batch_size=8)

Next, we define our model:

from transformers import AutoModelForSequenceClassification

model = AutoModelForSequenceClassification.from_pretrained("bert-base-cased", num_labels=2)

We are almost ready to write our training loop, the only two things are missing are an optimizer and a learning rate scheduler. The default optimizer used by the Trainer is AdamW:

from transformers import AdamW

optimizer = AdamW(model.parameters(), lr=5e-5)

Finally, the learning rate scheduler used by default is just a linear decay from the maximum value (5e-5 here) to 0:

from transformers import get_scheduler

num_epochs = 3
num_training_steps = num_epochs * len(train_dataloader)
lr_scheduler = get_scheduler(
    "linear",
    optimizer=optimizer,
    num_warmup_steps=0,
    num_training_steps=num_training_steps
)

One last thing, we will want to use the GPU if we have access to one (otherwise training might take several hours instead of a couple of minutes). To do this, we define a device we will put our model and our batches on.

import torch

device = torch.device("cuda") if torch.cuda.is_available() else torch.device("cpu")
model.to(device)

We now are ready to train! To get some sense of when it will be finished, we add a progress bar over our number of training steps, using the tqdm library.

from tqdm.auto import tqdm

progress_bar = tqdm(range(num_training_steps))

model.train()
for epoch in range(num_epochs):
    for batch in train_dataloader:
        batch = {k: v.to(device) for k, v in batch.items()}
        outputs = model(**batch)
        loss = outputs.loss
        loss.backward()

        optimizer.step()
        lr_scheduler.step()
        optimizer.zero_grad()
        progress_bar.update(1)

Note that if you are used to freezing the body of your pretrained model (like in computer vision) the above may seem a bit strange, as we are directly fine-tuning the whole model without taking any precaution. It actually works better this way for Transformers model (so this is not an oversight on our side). If you’re not familiar with what “freezing the body” of the model means, forget you read this paragraph.

Now to check the results, we need to write the evaluation loop. Like in the trainer section we will use a metric from the datasets library. Here we accumulate the predictions at each batch before computing the final result when the loop is finished.

metric= load_metric("accuracy")
model.eval()
for batch in eval_dataloader:
    batch = {k: v.to(device) for k, v in batch.items()}
    with torch.no_grad():
        outputs = model(**batch)

    logits = outputs.logits
    predictions = torch.argmax(logits, dim=-1)
    metric.add_batch(predictions=predictions, references=batch["labels"])

metric.compute()

Additional resources

To look at more fine-tuning examples you can refer to:

• 🤗 Transformers Examples which includes scripts to train on all common NLP tasks in PyTorch and TensorFlow.

• 🤗 Transformers Notebooks which contains various notebooks and in particular one per task (look for the how to finetune a model on xxx).