Donut-receipts-extract

Donut model was introduced in the paper OCR-free Document Understanding Transformer by Geewok et al. and first released in this repository.

=== V2 ===

This model has been retrained on an improved version of the AdamCodd/donut-receipts dataset (deduplicated, manually corrected). The new license for the V2 model is cc-by-nc-4.0. For commercial use rights, please contact me. Meanwhile, the V1 model remains available under the MIT license (under v1 branch).

It achieves the following results on the evaluation set:

• Loss: 0.326069
• Edit distance: 0.145293
• CER: 0.158358
• WER: 1.673989
• Mean accuracy: 0.895219
• F1: 0.977897

The task_prompt has been changed to <s_receipt> for the V2 (previously <s_cord-v2> for V1). Two new keys <s_svc> and <s_discount> have been added, <s_telephone> has been renamed to <s_phone>.

The V2 performs way better than the V1 as it has been trained on twice the resolution for the receipts, using a better dataset. Despite that, it's not perfect due to a lack of diverse receipts (the training dataset is still ~1100 receipts); for a future version, that will be the main focus.

=== V1 ====

This model is a finetune of the donut base model on the AdamCodd/donut-receipts dataset. Its purpose is to efficiently extract text from receipts.

It achieves the following results on the evaluation set:

• Loss: 0.498843
• Edit distance: 0.198315
• CER: 0.213929
• WER: 7.634032
• Mean accuracy: 0.843472

Model description

Donut consists of a vision encoder (Swin Transformer) and a text decoder (BART). Given an image, the encoder first encodes the image into a tensor of embeddings (of shape batch_size, seq_len, hidden_size), after which the decoder autoregressively generates text, conditioned on the encoding of the encoder.

How to use

import torch
import re
from PIL import Image
from transformers import DonutProcessor, VisionEncoderDecoderModel

device = torch.device('cuda:0' if torch.cuda.is_available() else 'cpu')
processor = DonutProcessor.from_pretrained("AdamCodd/donut-receipts-extract")
model = VisionEncoderDecoderModel.from_pretrained("AdamCodd/donut-receipts-extract")
model.to(device)

def load_and_preprocess_image(image_path: str, processor):
    """
    Load an image and preprocess it for the model.
    """
    image = Image.open(image_path).convert("RGB")
    pixel_values = processor(image, return_tensors="pt").pixel_values
    return pixel_values

def generate_text_from_image(model, image_path: str, processor, device):
    """
    Generate text from an image using the trained model.
    """
    # Load and preprocess the image
    pixel_values = load_and_preprocess_image(image_path, processor)
    pixel_values = pixel_values.to(device)

    # Generate output using model
    model.eval()
    with torch.no_grad():
        task_prompt = "<s_receipt>" # <s_cord-v2> for v1
        decoder_input_ids = processor.tokenizer(task_prompt, add_special_tokens=False, return_tensors="pt").input_ids
        decoder_input_ids = decoder_input_ids.to(device)
        generated_outputs = model.generate(
            pixel_values,
            decoder_input_ids=decoder_input_ids,
            max_length=model.decoder.config.max_position_embeddings, 
            pad_token_id=processor.tokenizer.pad_token_id,
            eos_token_id=processor.tokenizer.eos_token_id,
            early_stopping=True,
            bad_words_ids=[[processor.tokenizer.unk_token_id]],
            return_dict_in_generate=True
        )

    # Decode generated output
    decoded_text = processor.batch_decode(generated_outputs.sequences)[0]
    decoded_text = decoded_text.replace(processor.tokenizer.eos_token, "").replace(processor.tokenizer.pad_token, "")
    decoded_text = re.sub(r"<.*?>", "", decoded_text, count=1).strip()  # remove first task start token
    decoded_text = processor.token2json(decoded_text)
    return decoded_text

# Example usage
image_path = "path_to_your_image"  # Replace with your image path
extracted_text = generate_text_from_image(model, image_path, processor, device)
print("Extracted Text:", extracted_text)

Refer to the documentation for more code examples.

Intended uses & limitations

This fine-tuned model is specifically designed for extracting text from receipts and may not perform optimally on other types of documents. The dataset used is still suboptimal (numerous errors are still there) so this model will need to be retrained at a later date to improve its performance.

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 3e-05
• train_batch_size: 2
• eval_batch_size: 4
• seed: 42
• optimizer: AdamW with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 300
• num_epochs: 35
• weight_decay: 0.01

Framework versions

• Transformers 4.36.2
• Datasets 2.16.1
• Tokenizers 0.15.0
• Evaluate 0.4.1

If you want to support me, you can here.

BibTeX entry and citation info

@article{DBLP:journals/corr/abs-2111-15664,
  author    = {Geewook Kim and
               Teakgyu Hong and
               Moonbin Yim and
               Jinyoung Park and
               Jinyeong Yim and
               Wonseok Hwang and
               Sangdoo Yun and
               Dongyoon Han and
               Seunghyun Park},
  title     = {Donut: Document Understanding Transformer without {OCR}},
  journal   = {CoRR},
  volume    = {abs/2111.15664},
  year      = {2021},
  url       = {https://arxiv.org/abs/2111.15664},
  eprinttype = {arXiv},
  eprint    = {2111.15664},
  timestamp = {Thu, 02 Dec 2021 10:50:44 +0100},
  biburl    = {https://dblp.org/rec/journals/corr/abs-2111-15664.bib},
  bibsource = {dblp computer science bibliography, https://dblp.org}
}