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.

For an overview of the ecosystem of HuggingFace for computer vision (June 2022), refer to this notebook with corresponding video.

Currently, it contains the following demos:

• Audio Spectrogram Transformer (paper):
  • performing inference with ASTForAudioClassification to classify audio.
• BERT (paper):
  • fine-tuning BertForTokenClassification on a named entity recognition (NER) dataset.
  • fine-tuning BertForSequenceClassification for multi-label text classification.
• BEiT (paper):
  • understanding BeitForMaskedImageModeling
• CANINE (paper):
  • fine-tuning CanineForSequenceClassification on IMDb
• CLIPSeg (paper):
  • performing zero-shot image segmentation with CLIPSeg
• Conditional DETR (paper):
  • performing inference with ConditionalDetrForObjectDetection
  • fine-tuning ConditionalDetrForObjectDetection on a custom dataset (balloon)
• ConvNeXT (paper):
  • fine-tuning (and performing inference with) ConvNextForImageClassification
• DINO (paper):
  • visualize self-attention of Vision Transformers trained using the DINO method
• 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
• DPT (paper):
  • performing inference with DPT for monocular depth estimation
  • performing inference with DPT for semantic segmentation
• Deformable DETR (paper):
  • performing inference with DeformableDetrForObjectDetection
• DiT (paper):
  • performing inference with DiT for document image classification
• Donut (paper):
  • performing inference with Donut for document image classification
  • fine-tuning Donut for document image classification
  • performing inference with Donut for document visual question answering (DocVQA)
  • performing inference with Donut for document parsing
  • fine-tuning Donut for document parsing with PyTorch Lightning
• GIT (paper):
  • performing inference with GIT for image/video captioning and image/video question-answering
  • fine-tuning GIT on a custom image captioning dataset
• GLPN (paper):
  • performing inference with GLPNForDepthEstimation to illustrate monocular depth estimation
• GPT-J-6B (repository):
  • performing inference with GPTJForCausalLM to illustrate few-shot learning and code generation
• GroupViT (repository):
  • performing inference with GroupViTModel to illustrate zero-shot semantic segmentation
• ImageGPT (blog post):
  • (un)conditional image generation with ImageGPTForCausalLM
  • linear probing with ImageGPT
• LUKE (paper):
  • fine-tuning LukeForEntityPairClassification on a custom relation extraction dataset using PyTorch Lightning
• 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
• LayoutLMv3 (paper):
  • fine-tuning LayoutLMv3ForTokenClassification on the FUNSD dataset
• LayoutXLM (paper):
  • fine-tuning LayoutXLM on the XFUND benchmark for token classification
  • fine-tuning LayoutXLM on the XFUND benchmark for relation extraction
• MarkupLM (paper):
  • inference with MarkupLM to perform question answering on web pages
  • fine-tuning MarkupLMForTokenClassification on a toy dataset for NER on web pages
• Mask2Former (paper):
  • performing inference with Mask2Former for universal image segmentation:
• MaskFormer (paper):
  • performing inference with MaskFormer (both semantic and panoptic segmentation):
  • fine-tuning MaskFormer on a custom dataset for semantic segmentation
• OneFormer (paper):
  • performing inference with OneFormer for universal image segmentation:
• 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
• SAM (paper):
  • performing inference with MedSAM
  • fine-tuning SamModel on a custom dataset
• SegFormer (paper):
  • performing inference with SegformerForSemanticSegmentation
  • fine-tuning SegformerForSemanticSegmentation on custom data using native PyTorch
• 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
• Table Transformer (paper):
  • using the Table Transformer for table detection and table structure recognition
• 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
• UPerNet (paper):
  • performing inference with UperNetForSemanticSegmentation
• VideoMAE (paper):
  • performing inference with VideoMAEForVideoClassification
• ViLT (paper):
  • fine-tuning ViLT for visual question answering (VQA)
  • performing inference with ViLT to illustrate visual question answering (VQA)
  • masked language modeling (MLM) with a pre-trained ViLT model
  • performing inference with ViLT for image-text retrieval
  • performing inference with ViLT to illustrate natural language for visual reasoning (NLVR)
• ViTMAE (paper):
  • reconstructing pixel values with ViTMAEForPreTraining
• 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
• X-CLIP (paper):
  • performing zero-shot video classification with X-CLIP
  • zero-shot classifying a YouTube video with X-CLIP
• YOLOS (paper):
  • fine-tuning YolosForObjectDetection on a custom dataset
  • inference with YolosForObjectDetection

... 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
• DINO by Facebook 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
• MAE by Facebook AI
• ViLT by NAVER AI Lab
• ConvNeXT by Facebook AI
• DiT By Microsoft Research
• GLPN by KAIST
• DPT by Intel Labs
• YOLOS by School of EIC, Huazhong University of Science & Technology
• TAPEX by Microsoft Research
• LayoutLMv3 by Microsoft Research
• VideoMAE by Multimedia Computing Group, Nanjing University
• X-CLIP by Microsoft Research
• MarkupLM by Microsoft Research

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 creating a 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
from transformers import BertForSequenceClassification

# Instantiate pre-trained BERT model with randomly initialized classification head
model = BertForSequenceClassification.from_pretrained("bert-base-uncased")

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

# 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.).

Citation

Feel free to cite me when you use some of my tutorials :)

@misc{rogge2025transformerstutorials,
  author = {Rogge, Niels},
  title = {Tutorials},
  url = {[https://github.com/NielsRogge/tutorials](https://github.com/NielsRogge/Transformers-Tutorials)},
  year = {2025}
}