Vision Transformer(ViT) for Document Classification(DocLayNet)

This model is a fine-tuned Vision Transformer (ViT) for document layout classification based on the DocLayNet dataset.

Trained on images of the document categories from DocLayNet dataset where the categories namely(with their indexes) are :

{'financial_reports': 0,
 'government_tenders': 1,
 'laws_and_regulations': 2,
 'manuals': 3,
 'patents': 4,
 'scientific_articles': 5}

Model description

This model is built upon the google/vit-base-patch16-224-in21k Vision Transformer architecture and fine-tuned specifically for document layout classification. The base ViT model uses a patch size of 16x16 pixels and was pre-trained on ImageNet-21k. The model has been optimized to recognize and classify different types of document layouts from the DocLayNet dataset.

Training data

The model was trained on DocLayNet-base dataset, which is available on the Hugging Face Hub: pierreguillou/DocLayNet-base

DocLayNet is a comprehensive dataset for document layout analysis, containing various document types and their corresponding layout annotations.

Training procedure

Trained for 10 epochs on a single gpu for ~10 mins.

The training hyperparameters:

{
    'batch_size': 64,
    'num_epochs': 20,
    'learning_rate': 1e-4,
    'weight_decay': 0.05,
    'warmup_ratio': 0.2,
    'gradient_clip': 0.1,
    'dropout_rate': 0.1,
    'label_smoothing': 0.1,
    'optimizer': 'AdamW'
}

Evaluation results

The model achieved the following performance metrics on the test set:

Test Loss: 0.8622 Test Accuracy: 81.36%

Usage

from transformers import pipeline

# Load the model using the image-classification pipeline
pipe = pipeline("image-classification", model="kaixkhazaki/vit_doclaynet_base")

# Test it with an image
result = pipe("path_to_image.jpg")
print(result)

kaixkhazaki
/

vit_doclaynet_base