app.py

import torch
import streamlit as st

from PIL import Image
from transformers import VisionEncoderDecoderModel, VisionEncoderDecoderConfig , DonutProcessor


def run_prediction(sample):
    global pretrained_model, processor, task_prompt
    if isinstance(sample, dict):
        # prepare inputs
        pixel_values = torch.tensor(sample["pixel_values"]).unsqueeze(0)
    else:  # sample is an image
        # prepare encoder inputs
        pixel_values = processor(image, return_tensors="pt").pixel_values
    
    decoder_input_ids = processor.tokenizer(task_prompt, add_special_tokens=False, return_tensors="pt").input_ids

    # run inference
    outputs = pretrained_model.generate(
        pixel_values.to(device),
        decoder_input_ids=decoder_input_ids.to(device),
        max_length=pretrained_model.decoder.config.max_position_embeddings,
        early_stopping=True,
        pad_token_id=processor.tokenizer.pad_token_id,
        eos_token_id=processor.tokenizer.eos_token_id,
        use_cache=True,
        num_beams=1,
        bad_words_ids=[[processor.tokenizer.unk_token_id]],
        return_dict_in_generate=True,
    )

    # process output
    prediction = processor.batch_decode(outputs.sequences)[0]
    
    # post-processing
    if "cord" in task_prompt:
        prediction = prediction.replace(processor.tokenizer.eos_token, "").replace(processor.tokenizer.pad_token, "")
        prediction = re.sub(r"<.*?>", "", prediction, count=1).strip()  # remove first task start token
    prediction = processor.token2json(prediction)
    
    # load reference target
    if isinstance(sample, dict):
        target = processor.token2json(sample["target_sequence"])
    else:
        target = "<not_provided>"
    
    return prediction, target
    

task_prompt = f"<s>"

st.text('''
This is OCR-free Document Understanding Transformer nicknamed 🍩. It was fine-tuned with 1000 receipt images -> SROIE dataset.
The original 🍩 implementation can be found on: https://github.com/clovaai/donut
''')

with st.sidebar:
    information = st.radio(
    "What information inside the are you interested in?",
    ('Receipt Summary', 'Receipt Menu Details', 'Extract all!'))
    receipt = st.selectbox('Pick one receipt', ['1', '2', '3', '4', '5', '6'], index=5)

st.text(f'{information} mode is ON!\nTarget receipt: {receipt}\n(opening image @:./img/receipt-{receipt}.png)')

image = Image.open(f"./img/receipt-{receipt}.jpg")
st.image(image, caption='Your target receipt')

st.text(f'baking the 🍩...')
processor = DonutProcessor.from_pretrained("unstructuredio/donut-base-sroie")
pretrained_model = VisionEncoderDecoderModel.from_pretrained("unstructuredio/donut-base-sroie")

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

pretrained_model.encoder.to(torch.bfloat16)
pretrained_model.eval()

st.text(f'parsing receipt..')
parsed_receipt_info = run_prediction(image)
st.text(f'\nRaw output:\n{parsed_receipt_info}')