Transformers-Tutorials

Hi there!

This repository contains demos I made with the Transformers library by 🤗 HuggingFace. Currently, all of them are implemented in PyTorch.

NOTE: if you are not familiar with HuggingFace and/or Transformers, I highly recommend to check out our free course, which introduces you to several Transformer architectures (such as BERT, GPT-2, T5, BART, etc.), as well as an overview of the HuggingFace libraries, including Transformers, Tokenizers, Datasets, Accelerate and the hub.

Currently, it contains the following demos:

• BERT (paper):
  • fine-tuning BertForTokenClassification on a named entity recognition (NER) dataset.
  • fine-tuning BertForSequenceClassification for multi-label text classification.
• CANINE (paper):
  • fine-tuning CanineForSequenceClassification on IMDb
• DETR (paper):
  • performing inference with DetrForObjectDetection
  • fine-tuning DetrForObjectDetection on a custom object detection dataset
  • evaluating DetrForObjectDetection on the COCO detection 2017 validation set
  • performing inference with DetrForSegmentation
  • fine-tuning DetrForSegmentation on COCO panoptic 2017
• GPT-J-6B (repository):
  • performing inference with GPTJForCausalLM to illustrate few-shot learning and code generation
• ImageGPT (blog post):
  • (un)conditional image generation with ImageGPTForCausalLM
  • linear probing with ImageGPT
• LayoutLM (paper):
  • fine-tuning LayoutLMForTokenClassification on the FUNSD dataset
  • fine-tuning LayoutLMForSequenceClassification on the RVL-CDIP dataset
  • adding image embeddings to LayoutLM during fine-tuning on the FUNSD dataset
• LayoutLMv2 (paper):
  • fine-tuning LayoutLMv2ForSequenceClassification on RVL-CDIP
  • fine-tuning LayoutLMv2ForTokenClassification on FUNSD
  • fine-tuning LayoutLMv2ForTokenClassification on FUNSD using the 🤗 Trainer
  • performing inference with LayoutLMv2ForTokenClassification on FUNSD
  • true inference with LayoutLMv2ForTokenClassification (when no labels are available) + Gradio demo
  • fine-tuning LayoutLMv2ForTokenClassification on CORD
  • fine-tuning LayoutLMv2ForQuestionAnswering on DOCVQA
• LUKE (paper):
  • fine-tuning LukeForEntityPairClassification on a custom relation extraction dataset using PyTorch Lightning
• SegFormer (paper):
  • performing inference with SegformerForSemanticSegmentation
  • fine-tuning SegformerForSemanticSegmentation on custom data using native PyTorch
• Perceiver IO (paper):
  • showcasing masked language modeling and image classification with the Perceiver
  • fine-tuning the Perceiver for image classification
  • fine-tuning the Perceiver for text classification
  • predicting optical flow between a pair of images with PerceiverForOpticalFlow
  • auto-encoding a video (images, audio, labels) with PerceiverForMultimodalAutoencoding
• T5 (paper):
  • fine-tuning T5ForConditionalGeneration on a Dutch summarization dataset on TPU using HuggingFace Accelerate
  • fine-tuning T5ForConditionalGeneration (CodeT5) for Ruby code summarization using PyTorch Lightning
• TAPAS (paper):
  • fine-tuning TapasForQuestionAnswering on the Microsoft Sequential Question Answering (SQA) dataset
  • evaluating TapasForSequenceClassification on the Table Fact Checking (TabFact) dataset
• TrOCR (paper):
  • performing inference with TrOCR to illustrate optical character recognition with Transformers, as well as making a Gradio demo
  • fine-tuning TrOCR on the IAM dataset using the Seq2SeqTrainer
  • fine-tuning TrOCR on the IAM dataset using native PyTorch
  • evaluating TrOCR on the IAM test set
• Vision Transformer (paper):
  • performing inference with ViTForImageClassification
  • fine-tuning ViTForImageClassification on CIFAR-10 using PyTorch Lightning
  • fine-tuning ViTForImageClassification on CIFAR-10 using the 🤗 Trainer

... more to come! 🤗

If you have any questions regarding these demos, feel free to open an issue on this repository.

Btw, I was also the main contributor to add the following algorithms to the library:

• TAbular PArSing (TAPAS) by Google AI
• Vision Transformer (ViT) by Google AI
• Data-efficient Image Transformers (DeiT) by Facebook AI
• LUKE by Studio Ousia
• DEtection TRansformers (DETR) by Facebook AI
• CANINE by Google AI
• BEiT by Microsoft Research
• LayoutLMv2 (and LayoutXLM) by Microsoft Research
• TrOCR by Microsoft Research
• SegFormer by NVIDIA
• ImageGPT by OpenAI
• Perceiver by Deepmind

All of them were an incredible learning experience. I can recommend anyone to contribute an AI algorithm to the library!

Data preprocessing

Regarding preparing your data for a PyTorch model, there are a few options:

• a native PyTorch dataset + dataloader. This is the standard way to prepare data for a PyTorch model, namely by subclassing torch.utils.data.Dataset, and then a creating corresponding DataLoader (which is a Python generator that allows to loop over the items of a dataset). When subclassing the Dataset class, one needs to implement 3 methods: __init__, __len__ (which returns the number of examples of the dataset) and __getitem__ (which returns an example of the dataset, given an integer index). Here's an example of creating a basic text classification dataset (assuming one has a CSV that contains 2 columns, namely "text" and "label"):

from torch.utils.data import Dataset

class CustomTrainDataset(Dataset):
    def __init__(self, df, tokenizer):
        self.df = df
        self.tokenizer = tokenizer

    def __len__(self):
        return len(self.df)

    def __getitem__(self, idx):
        # get item
        item = df.iloc[idx]
        text = item['text']
        label = item['label']
        # encode text
        encoding = self.tokenizer(text, padding="max_length", max_length=128, truncation=True, return_tensors="pt")
        # remove batch dimension which the tokenizer automatically adds
        encoding = {k:v.squeeze() for k,v in encoding.items()}
        # add label
        encoding["label"] = torch.tensor(label)
        
        return encoding

Instantiating the dataset then happens as follows:

from transformers import BertTokenizer
import pandas as pd

tokenizer = BertTokenizer.from_pretrained("bert-base-uncased")
df = pd.read_csv("path_to_your_csv")

train_dataset = CustomTrainDataset(df=df tokenizer=tokenizer)

Accessing the first example of the dataset can then be done as follows:

encoding = train_dataset[0]

In practice, one creates a corresponding DataLoader, that allows to get batches from the dataset:

from torch.utils.data import DataLoader

train_dataloader = DataLoader(train_dataset, batch_size=4, shuffle=True)

I often check whether the data is created correctly by fetching the first batch from the data loader, and then printing out the shapes of the tensors, decoding the input_ids back to text, etc.

batch = next(iter(train_dataloader))
for k,v in batch.items():
    print(k, v.shape)
# decode the input_ids of the first example of the batch
print(tokenizer.decode(batch['input_ids'][0].tolist())

• HuggingFace Datasets. Datasets is a library by HuggingFace that allows to easily load and process data in a very fast and memory-efficient way. It is backed by Apache Arrow, and has cool features such as memory-mapping, which allow you to only load data into RAM when it is required. It only has deep interoperability with the HuggingFace hub, allowing to easily load well-known datasets as well as share your own with the community.

Loading a custom dataset as a Dataset object can be done as follows (you can install datasets using pip install datasets):

from datasets import load_dataset

dataset = load_dataset('csv', data_files={'train': ['my_train_file_1.csv', 'my_train_file_2.csv'] 'test': 'my_test_file.csv'})

Here I'm loading local CSV files, but there are other formats supported (including JSON, Parquet, txt) as well as loading data from a local Pandas dataframe or dictionary for instance. You can check out the docs for all details.

Training frameworks

Regarding fine-tuning Transformer models (or more generally, PyTorch models), there are a few options:

• using native PyTorch. This is the most basic way to train a model, and requires the user to manually write the training loop. The advantage is that this is very easy to debug. The disadvantage is that one needs to implement training him/herself, such as setting the model in the appropriate mode (model.train()/model.eval()), handle device placement (model.to(device)), etc. A typical training loop in PyTorch looks as follows (inspired by this great PyTorch intro tutorial):

import torch

model = ...

# I almost always use a learning rate of 5e-5 when fine-tuning Transformer based models
optimizer = torch.optim.Adam(model.parameters(), lr=5-e5)

# put model on GPU, if available
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
model.to(device)

for epoch in range(epochs):
    model.train()
    train_loss = 0.0
    for batch in train_dataloader:
        # put batch on device
        batch = {k:v.to(device) for k,v in batch.items()}
        
        # forward pass
        outputs = model(**batch)
        loss = outputs.loss
        
        train_loss += loss.item()
        
        loss.backward()
        optimizer.step()
        optimizer.zero_grad()

    print("Loss after epoch {epoch}:", train_loss/len(train_dataloader))
    
    model.eval()
    val_loss = 0.0
    with torch.no_grad():
        for batch in eval_dataloader:
            # put batch on device
            batch = {k:v.to(device) for k,v in batch.items()}
            
            # forward pass
            outputs = model(**batch)
            loss = outputs.logits
            
            val_loss += loss.item()
                  
    print("Validation loss after epoch {epoch}:", val_loss/len(eval_dataloader))

• PyTorch Lightning (PL). PyTorch Lightning is a framework that automates the training loop written above, by abstracting it away in a Trainer object. Users don't need to write the training loop themselves anymore, instead they can just do trainer = Trainer() and then trainer.fit(model). The advantage is that you can start training models very quickly (hence the name lightning), as all training-related code is handled by the Trainer object. The disadvantage is that it may be more difficult to debug your model, as the training and evaluation is now abstracted away.
• HuggingFace Trainer. The HuggingFace Trainer API can be seen as a framework similar to PyTorch Lightning in the sense that it also abstracts the training away using a Trainer object. However, contrary to PyTorch Lightning, it is not meant not be a general framework. Rather, it is made especially for fine-tuning Transformer-based models available in the HuggingFace Transformers library. The Trainer also has an extension called Seq2SeqTrainer for encoder-decoder models, such as BART, T5 and the EncoderDecoderModel classes. Note that all PyTorch example scripts of the Transformers library make use of the Trainer.
• HuggingFace Accelerate: Accelerate is a new project, that is made for people who still want to write their own training loop (as shown above), but would like to make it work automatically irregardless of the hardware (i.e. multiple GPUs, TPU pods, mixed precision, etc.).