Transformers
  • 🌍GET STARTED
    • Transformers
    • Quick tour
    • Installation
  • 🌍TUTORIALS
    • Run inference with pipelines
    • Write portable code with AutoClass
    • Preprocess data
    • Fine-tune a pretrained model
    • Train with a script
    • Set up distributed training with BOINC AI Accelerate
    • Load and train adapters with BOINC AI PEFT
    • Share your model
    • Agents
    • Generation with LLMs
  • 🌍TASK GUIDES
    • 🌍NATURAL LANGUAGE PROCESSING
      • Text classification
      • Token classification
      • Question answering
      • Causal language modeling
      • Masked language modeling
      • Translation
      • Summarization
      • Multiple choice
    • 🌍AUDIO
      • Audio classification
      • Automatic speech recognition
    • 🌍COMPUTER VISION
      • Image classification
      • Semantic segmentation
      • Video classification
      • Object detection
      • Zero-shot object detection
      • Zero-shot image classification
      • Depth estimation
    • 🌍MULTIMODAL
      • Image captioning
      • Document Question Answering
      • Visual Question Answering
      • Text to speech
    • 🌍GENERATION
      • Customize the generation strategy
    • 🌍PROMPTING
      • Image tasks with IDEFICS
  • 🌍DEVELOPER GUIDES
    • Use fast tokenizers from BOINC AI Tokenizers
    • Run inference with multilingual models
    • Use model-specific APIs
    • Share a custom model
    • Templates for chat models
    • Run training on Amazon SageMaker
    • Export to ONNX
    • Export to TFLite
    • Export to TorchScript
    • Benchmarks
    • Notebooks with examples
    • Community resources
    • Custom Tools and Prompts
    • Troubleshoot
  • 🌍PERFORMANCE AND SCALABILITY
    • Overview
    • 🌍EFFICIENT TRAINING TECHNIQUES
      • Methods and tools for efficient training on a single GPU
      • Multiple GPUs and parallelism
      • Efficient training on CPU
      • Distributed CPU training
      • Training on TPUs
      • Training on TPU with TensorFlow
      • Training on Specialized Hardware
      • Custom hardware for training
      • Hyperparameter Search using Trainer API
    • 🌍OPTIMIZING INFERENCE
      • Inference on CPU
      • Inference on one GPU
      • Inference on many GPUs
      • Inference on Specialized Hardware
    • Instantiating a big model
    • Troubleshooting
    • XLA Integration for TensorFlow Models
    • Optimize inference using `torch.compile()`
  • 🌍CONTRIBUTE
    • How to contribute to transformers?
    • How to add a model to BOINC AI Transformers?
    • How to convert a BOINC AI Transformers model to TensorFlow?
    • How to add a pipeline to BOINC AI Transformers?
    • Testing
    • Checks on a Pull Request
  • 🌍CONCEPTUAL GUIDES
    • Philosophy
    • Glossary
    • What BOINC AI Transformers can do
    • How BOINC AI Transformers solve tasks
    • The Transformer model family
    • Summary of the tokenizers
    • Attention mechanisms
    • Padding and truncation
    • BERTology
    • Perplexity of fixed-length models
    • Pipelines for webserver inference
    • Model training anatomy
  • 🌍API
    • 🌍MAIN CLASSES
      • Agents and Tools
      • 🌍Auto Classes
        • Extending the Auto Classes
        • AutoConfig
        • AutoTokenizer
        • AutoFeatureExtractor
        • AutoImageProcessor
        • AutoProcessor
        • Generic model classes
          • AutoModel
          • TFAutoModel
          • FlaxAutoModel
        • Generic pretraining classes
          • AutoModelForPreTraining
          • TFAutoModelForPreTraining
          • FlaxAutoModelForPreTraining
        • Natural Language Processing
          • AutoModelForCausalLM
          • TFAutoModelForCausalLM
          • FlaxAutoModelForCausalLM
          • AutoModelForMaskedLM
          • TFAutoModelForMaskedLM
          • FlaxAutoModelForMaskedLM
          • AutoModelForMaskGenerationge
          • TFAutoModelForMaskGeneration
          • AutoModelForSeq2SeqLM
          • TFAutoModelForSeq2SeqLM
          • FlaxAutoModelForSeq2SeqLM
          • AutoModelForSequenceClassification
          • TFAutoModelForSequenceClassification
          • FlaxAutoModelForSequenceClassification
          • AutoModelForMultipleChoice
          • TFAutoModelForMultipleChoice
          • FlaxAutoModelForMultipleChoice
          • AutoModelForNextSentencePrediction
          • TFAutoModelForNextSentencePrediction
          • FlaxAutoModelForNextSentencePrediction
          • AutoModelForTokenClassification
          • TFAutoModelForTokenClassification
          • FlaxAutoModelForTokenClassification
          • AutoModelForQuestionAnswering
          • TFAutoModelForQuestionAnswering
          • FlaxAutoModelForQuestionAnswering
          • AutoModelForTextEncoding
          • TFAutoModelForTextEncoding
        • Computer vision
          • AutoModelForDepthEstimation
          • AutoModelForImageClassification
          • TFAutoModelForImageClassification
          • FlaxAutoModelForImageClassification
          • AutoModelForVideoClassification
          • AutoModelForMaskedImageModeling
          • TFAutoModelForMaskedImageModeling
          • AutoModelForObjectDetection
          • AutoModelForImageSegmentation
          • AutoModelForImageToImage
          • AutoModelForSemanticSegmentation
          • TFAutoModelForSemanticSegmentation
          • AutoModelForInstanceSegmentation
          • AutoModelForUniversalSegmentation
          • AutoModelForZeroShotImageClassification
          • TFAutoModelForZeroShotImageClassification
          • AutoModelForZeroShotObjectDetection
        • Audio
          • AutoModelForAudioClassification
          • AutoModelForAudioFrameClassification
          • TFAutoModelForAudioFrameClassification
          • AutoModelForCTC
          • AutoModelForSpeechSeq2Seq
          • TFAutoModelForSpeechSeq2Seq
          • FlaxAutoModelForSpeechSeq2Seq
          • AutoModelForAudioXVector
          • AutoModelForTextToSpectrogram
          • AutoModelForTextToWaveform
        • Multimodal
          • AutoModelForTableQuestionAnswering
          • TFAutoModelForTableQuestionAnswering
          • AutoModelForDocumentQuestionAnswering
          • TFAutoModelForDocumentQuestionAnswering
          • AutoModelForVisualQuestionAnswering
          • AutoModelForVision2Seq
          • TFAutoModelForVision2Seq
          • FlaxAutoModelForVision2Seq
      • Callbacks
      • Configuration
      • Data Collator
      • Keras callbacks
      • Logging
      • Models
      • Text Generation
      • ONNX
      • Optimization
      • Model outputs
      • Pipelines
      • Processors
      • Quantization
      • Tokenizer
      • Trainer
      • DeepSpeed Integration
      • Feature Extractor
      • Image Processor
    • 🌍MODELS
      • 🌍TEXT MODELS
        • ALBERT
        • BART
        • BARThez
        • BARTpho
        • BERT
        • BertGeneration
        • BertJapanese
        • Bertweet
        • BigBird
        • BigBirdPegasus
        • BioGpt
        • Blenderbot
        • Blenderbot Small
        • BLOOM
        • BORT
        • ByT5
        • CamemBERT
        • CANINE
        • CodeGen
        • CodeLlama
        • ConvBERT
        • CPM
        • CPMANT
        • CTRL
        • DeBERTa
        • DeBERTa-v2
        • DialoGPT
        • DistilBERT
        • DPR
        • ELECTRA
        • Encoder Decoder Models
        • ERNIE
        • ErnieM
        • ESM
        • Falcon
        • FLAN-T5
        • FLAN-UL2
        • FlauBERT
        • FNet
        • FSMT
        • Funnel Transformer
        • GPT
        • GPT Neo
        • GPT NeoX
        • GPT NeoX Japanese
        • GPT-J
        • GPT2
        • GPTBigCode
        • GPTSAN Japanese
        • GPTSw3
        • HerBERT
        • I-BERT
        • Jukebox
        • LED
        • LLaMA
        • LLama2
        • Longformer
        • LongT5
        • LUKE
        • M2M100
        • MarianMT
        • MarkupLM
        • MBart and MBart-50
        • MEGA
        • MegatronBERT
        • MegatronGPT2
        • Mistral
        • mLUKE
        • MobileBERT
        • MPNet
        • MPT
        • MRA
        • MT5
        • MVP
        • NEZHA
        • NLLB
        • NLLB-MoE
        • Nyströmformer
        • Open-Llama
        • OPT
        • Pegasus
        • PEGASUS-X
        • Persimmon
        • PhoBERT
        • PLBart
        • ProphetNet
        • QDQBert
        • RAG
        • REALM
        • Reformer
        • RemBERT
        • RetriBERT
        • RoBERTa
        • RoBERTa-PreLayerNorm
        • RoCBert
        • RoFormer
        • RWKV
        • Splinter
        • SqueezeBERT
        • SwitchTransformers
        • T5
        • T5v1.1
        • TAPEX
        • Transformer XL
        • UL2
        • UMT5
        • X-MOD
        • XGLM
        • XLM
        • XLM-ProphetNet
        • XLM-RoBERTa
        • XLM-RoBERTa-XL
        • XLM-V
        • XLNet
        • YOSO
      • 🌍VISION MODELS
        • BEiT
        • BiT
        • Conditional DETR
        • ConvNeXT
        • ConvNeXTV2
        • CvT
        • Deformable DETR
        • DeiT
        • DETA
        • DETR
        • DiNAT
        • DINO V2
        • DiT
        • DPT
        • EfficientFormer
        • EfficientNet
        • FocalNet
        • GLPN
        • ImageGPT
        • LeViT
        • Mask2Former
        • MaskFormer
        • MobileNetV1
        • MobileNetV2
        • MobileViT
        • MobileViTV2
        • NAT
        • PoolFormer
        • Pyramid Vision Transformer (PVT)
        • RegNet
        • ResNet
        • SegFormer
        • SwiftFormer
        • Swin Transformer
        • Swin Transformer V2
        • Swin2SR
        • Table Transformer
        • TimeSformer
        • UperNet
        • VAN
        • VideoMAE
        • Vision Transformer (ViT)
        • ViT Hybrid
        • ViTDet
        • ViTMAE
        • ViTMatte
        • ViTMSN
        • ViViT
        • YOLOS
      • 🌍AUDIO MODELS
        • Audio Spectrogram Transformer
        • Bark
        • CLAP
        • EnCodec
        • Hubert
        • MCTCT
        • MMS
        • MusicGen
        • Pop2Piano
        • SEW
        • SEW-D
        • Speech2Text
        • Speech2Text2
        • SpeechT5
        • UniSpeech
        • UniSpeech-SAT
        • VITS
        • Wav2Vec2
        • Wav2Vec2-Conformer
        • Wav2Vec2Phoneme
        • WavLM
        • Whisper
        • XLS-R
        • XLSR-Wav2Vec2
      • 🌍MULTIMODAL MODELS
        • ALIGN
        • AltCLIP
        • BLIP
        • BLIP-2
        • BridgeTower
        • BROS
        • Chinese-CLIP
        • CLIP
        • CLIPSeg
        • Data2Vec
        • DePlot
        • Donut
        • FLAVA
        • GIT
        • GroupViT
        • IDEFICS
        • InstructBLIP
        • LayoutLM
        • LayoutLMV2
        • LayoutLMV3
        • LayoutXLM
        • LiLT
        • LXMERT
        • MatCha
        • MGP-STR
        • Nougat
        • OneFormer
        • OWL-ViT
        • Perceiver
        • Pix2Struct
        • Segment Anything
        • Speech Encoder Decoder Models
        • TAPAS
        • TrOCR
        • TVLT
        • ViLT
        • Vision Encoder Decoder Models
        • Vision Text Dual Encoder
        • VisualBERT
        • X-CLIP
      • 🌍REINFORCEMENT LEARNING MODELS
        • Decision Transformer
        • Trajectory Transformer
      • 🌍TIME SERIES MODELS
        • Autoformer
        • Informer
        • Time Series Transformer
      • 🌍GRAPH MODELS
        • Graphormer
  • 🌍INTERNAL HELPERS
    • Custom Layers and Utilities
    • Utilities for pipelines
    • Utilities for Tokenizers
    • Utilities for Trainer
    • Utilities for Generation
    • Utilities for Image Processors
    • Utilities for Audio processing
    • General Utilities
    • Utilities for Time Series
Powered by GitBook
On this page
  • Prepare a dataset
  • Prepare a dataset
  • Train with PyTorch Trainer
  • Train a TensorFlow model with Keras
  • Train in native PyTorch
  • Additional resources
  1. TUTORIALS

Fine-tune a pretrained model

PreviousPreprocess dataNextTrain with a script

Last updated 1 year ago

There are significant benefits to using a pretrained model. It reduces computation costs, your carbon footprint, and allows you to use state-of-the-art models without having to train one from scratch. 🌎 Transformers provides access to thousands of pretrained models for a wide range of tasks. When you use a pretrained model, you train it on a dataset specific to your task. This is known as fine-tuning, an incredibly powerful training technique. In this tutorial, you will fine-tune a pretrained model with a deep learning framework of your choice:

  • Fine-tune a pretrained model with🌎 Transformers .

  • Fine-tune a pretrained model in TensorFlow with Keras.

  • Fine-tune a pretrained model in native PyTorch.

Prepare a dataset

There are significant benefits to using a pretrained model. It reduces computation costs, your carbon footprint, and allows you to use state-of-the-art models without having to train one from scratch. 🌎 Transformers provides access to thousands of pretrained models for a wide range of tasks. When you use a pretrained model, you train it on a dataset specific to your task. This is known as fine-tuning, an incredibly powerful training technique. In this tutorial, you will fine-tune a pretrained model with a deep learning framework of your choice:

  • Fine-tune a pretrained model with 🌎 Transformers .

  • Fine-tune a pretrained model in TensorFlow with Keras.

  • Fine-tune a pretrained model in native PyTorch.

Prepare a dataset

Train with PyTorch Trainer

🌎 Transformers provides a class optimized for training🌎 Transformers models, making it easier to start training without manually writing your own training loop. The API supports a wide range of training options and features such as logging, gradient accumulation, and mixed precision.

Start by loading your model and specify the number of expected labels. From the Yelp Review , you know there are five labels:

Copied

>>> from transformers import AutoModelForSequenceClassification

>>> model = AutoModelForSequenceClassification.from_pretrained("bert-base-cased", num_labels=5)

You will see a warning about some of the pretrained weights not being used and some weights being randomly initialized. Don’t worry, this is completely normal! The pretrained head of the BERT model is discarded, and replaced with a randomly initialized classification head. You will fine-tune this new model head on your sequence classification task, transferring the knowledge of the pretrained model to it.

Training hyperparameters

Specify where to save the checkpoints from your training:

Copied

>>> from transformers import TrainingArguments

>>> training_args = TrainingArguments(output_dir="test_trainer")

Evaluate

Copied

>>> import numpy as np
>>> import evaluate

>>> metric = evaluate.load("accuracy")

Call compute on metric to calculate the accuracy of your predictions. Before passing your predictions to compute, you need to convert the predictions to logits (remember all🌎Transformers models return logits):

Copied

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

If you’d like to monitor your evaluation metrics during fine-tuning, specify the evaluation_strategy parameter in your training arguments to report the evaluation metric at the end of each epoch:

Copied

>>> from transformers import TrainingArguments, Trainer

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

Trainer

Copied

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

Copied

>>> trainer.train()

Train a TensorFlow model with Keras

You can also train 🌎 Transformers models in TensorFlow with the Keras API!

Loading data for Keras

When you want to train a 🌎 Transformers model with the Keras API, you need to convert your dataset to a format that Keras understands. If your dataset is small, you can just convert the whole thing to NumPy arrays and pass it to Keras. Let’s try that first before we do anything more complicated.

Copied

from datasets import load_dataset

dataset = load_dataset("glue", "cola")
dataset = dataset["train"]  # Just take the training split for now

Next, load a tokenizer and tokenize the data as NumPy arrays. Note that the labels are already a list of 0 and 1s, so we can just convert that directly to a NumPy array without tokenization!

Copied

from transformers import AutoTokenizer

tokenizer = AutoTokenizer.from_pretrained("bert-base-cased")
tokenized_data = tokenizer(dataset["sentence"], return_tensors="np", padding=True)
# Tokenizer returns a BatchEncoding, but we convert that to a dict for Keras
tokenized_data = dict(tokenized_data)

labels = np.array(dataset["label"])  # Label is already an array of 0 and 1

Copied

from transformers import TFAutoModelForSequenceClassification
from tensorflow.keras.optimizers import Adam

# Load and compile our model
model = TFAutoModelForSequenceClassification.from_pretrained("bert-base-cased")
# Lower learning rates are often better for fine-tuning transformers
model.compile(optimizer=Adam(3e-5))  # No loss argument!

model.fit(tokenized_data, labels)

You don’t have to pass a loss argument to your models when you compile() them! BOINC AI models automatically choose a loss that is appropriate for their task and model architecture if this argument is left blank. You can always override this by specifying a loss yourself if you want to!

This approach works great for smaller datasets, but for larger datasets, you might find it starts to become a problem. Why? Because the tokenized array and labels would have to be fully loaded into memory, and because NumPy doesn’t handle “jagged” arrays, so every tokenized sample would have to be padded to the length of the longest sample in the whole dataset. That’s going to make your array even bigger, and all those padding tokens will slow down training too!

Loading data as a tf.data.Dataset

If you want to avoid slowing down training, you can load your data as a tf.data.Dataset instead. Although you can write your own tf.data pipeline if you want, we have two convenience methods for doing this:

Copied

def tokenize_dataset(data):
    # Keys of the returned dictionary will be added to the dataset as columns
    return tokenizer(data["text"])


dataset = dataset.map(tokenize_dataset)

Remember that BOINC AI datasets are stored on disk by default, so this will not inflate your memory usage! Once the columns have been added, you can stream batches from the dataset and add padding to each batch, which greatly reduces the number of padding tokens compared to padding the entire dataset.

Copied

>>> tf_dataset = model.prepare_tf_dataset(dataset["train"], batch_size=16, shuffle=True, tokenizer=tokenizer)

Once you’ve created a tf.data.Dataset, you can compile and fit the model as before:

Copied

model.compile(optimizer=Adam(3e-5))  # No loss argument!

model.fit(tf_dataset)

Train in native PyTorch

At this point, you may need to restart your notebook or execute the following code to free some memory:

Copied

del model
del trainer
torch.cuda.empty_cache()

Next, manually postprocess tokenized_dataset to prepare it for training.

  1. Remove the text column because the model does not accept raw text as an input:

    Copied

    >>> tokenized_datasets = tokenized_datasets.remove_columns(["text"])
  2. Rename the label column to labels because the model expects the argument to be named labels:

    Copied

    >>> tokenized_datasets = tokenized_datasets.rename_column("label", "labels")
  3. Set the format of the dataset to return PyTorch tensors instead of lists:

    Copied

    >>> tokenized_datasets.set_format("torch")

Then create a smaller subset of the dataset as previously shown to speed up the fine-tuning:

Copied

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

DataLoader

Create a DataLoader for your training and test datasets so you can iterate over batches of data:

Copied

>>> 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)

Load your model with the number of expected labels:

Copied

>>> from transformers import AutoModelForSequenceClassification

>>> model = AutoModelForSequenceClassification.from_pretrained("bert-base-cased", num_labels=5)

Optimizer and learning rate scheduler

Copied

>>> from torch.optim import AdamW

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

Copied

>>> from transformers import get_scheduler

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

Lastly, specify device to use a GPU if you have access to one. Otherwise, training on a CPU may take several hours instead of a couple of minutes.

Copied

>>> import torch

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

Great, now you are ready to train! 🌎

Training loop

Copied

>>> 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)

Evaluate

Copied

>>> import evaluate

>>> metric = evaluate.load("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

For more fine-tuning examples, refer to:

Next, create a class which contains all the hyperparameters you can tune as well as flags for activating different training options. For this tutorial you can start with the default training , but feel free to experiment with these to find your optimal settings.

does not automatically evaluate model performance during training. You’ll need to pass a function to compute and report metrics. The 🌎 library provides a simple function you can load with the evaluate.load (see this for more information) function:

Create a object with your model, training arguments, training and test datasets, and evaluation function:

Then fine-tune your model by calling :

First, load a dataset. We’ll use the CoLA dataset from the , since it’s a simple binary text classification task, and just take the training split for now.

Finally, load, , and the model. Note that Transformers models all have a default task-relevant loss function, so you don’t need to specify one unless you want to:

: This is the method we recommend in most cases. Because it is a method on your model, it can inspect the model to automatically figure out which columns are usable as model inputs, and discard the others to make a simpler, more performant dataset.

: This method is more low-level, and is useful when you want to exactly control how your dataset is created, by specifying exactly which columns and label_cols to include.

Before you can use , you will need to add the tokenizer outputs to your dataset as columns, as shown in the following code sample:

Note that in the code sample above, you need to pass the tokenizer to prepare_tf_dataset so it can correctly pad batches as they’re loaded. If all the samples in your dataset are the same length and no padding is necessary, you can skip this argument. If you need to do something more complex than just padding samples (e.g. corrupting tokens for masked language modelling), you can use the collate_fn argument instead to pass a function that will be called to transform the list of samples into a batch and apply any preprocessing you want. See our or to see this approach in action.

takes care of the training loop and allows you to fine-tune a model in a single line of code. For users who prefer to write their own training loop, you can also fine-tune a 🌎Transformers model in native PyTorch.

Create an optimizer and learning rate scheduler to fine-tune the model. Let’s use the optimizer from PyTorch:

Create the default learning rate scheduler from :

Get free access to a cloud GPU if you don’t have one with a hosted notebook like or .

To keep track of your training progress, use the library to add a progress bar over the number of training steps:

Just like how you added an evaluation function to , you need to do the same when you write your own training loop. But instead of calculating and reporting the metric at the end of each epoch, this time you’ll accumulate all the batches with add_batch and calculate the metric at the very end.

🌎 includes scripts to train common NLP tasks in PyTorch and TensorFlow.

🌎 contains various notebooks on how to fine-tune a model for specific tasks in PyTorch and TensorFlow.

🌍
Trainer
Trainer
Trainer
Trainer
dataset card
TrainingArguments
hyperparameters
Trainer
Trainer
Evaluate
accuracy
quicktour
Trainer
train()
GLUE benchmark
compile
fit
prepare_tf_dataset()
to_tf_dataset
prepare_tf_dataset()
examples
notebooks
Trainer
AdamW
Trainer
Colaboratory
SageMaker StudioLab
tqdm
Trainer
Transformers Examples
Transformers Notebooks