app.py

import torch
import streamlit as st

from PIL import Image
from io import BytesIO
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. set_page_config(layout="wide")

# logo = Image.open("./img/rsz_unstructured_logo.png")
# st.image(logo)

st.markdown('''
### Donut Common Crawl
Experimental OCR-free Document Understanding Vision Transformer nicknamed 🍩, fine-tuned with few samples of the common-crawl with some specific document elements.
''')

with st.sidebar:
    information = st.radio(
    "Choose one predictor:?",
    ('Base Common-Crawl 🍩', 'Hierarchical Common-Crawl 🍩'))
    image_choice = st.selectbox('Pick one 📑', ['1', '2', '3', '4', '5'], index=1)
        
st.text(f'{information} mode is ON!\nTarget 📑: {image_choice}')  # \n(opening image @:./img/receipt-{receipt}.png)')

col1, col2 = st.columns(2)

image_choice_map = {
    '1': 'commoncrawl_amandalacombznewspolice-bust-man-sawed-oal_1.jpg',
    '2': 'commoncrawl_canyonhillschroniclecomtagwomens-basketbll_0.png',
    '3': 'commoncrawl_celstuttgartdeideaa-different-stort-of-nfe_0.png',
    '4': 'commoncrawl_abusoftbplacednetsearchphpsearchidunanswae_0.png',
    '5': 'commoncrawl_addxorrolblogspotcomessays-about-managemml_1.png'
}
image = Image.open(f'samples/{image_choice_map[image_choice]}')
with col1:
    st.image(image, caption='Your target sample')

if st.button('Parse sample! 🐍'):
    image = image.convert('RGB')
    image.save('./target_image.jpg')
    image = Image.open('./target_image.jpg')
    with st.spinner(f'baking the 🍩s...'):
        if information == 'Base Common-Crawl 🍩':
            processor = DonutProcessor.from_pretrained("laverdes/donut-web")  # laverdes/donut-commoncrawl
            pretrained_model = VisionEncoderDecoderModel.from_pretrained("laverdes/donut-web")   # laverdes/donut-commoncrawl
            task_prompt = f"<s>"
            device = "cuda" if torch.cuda.is_available() else "cpu"
            pretrained_model.to(device)
        
        elif information == 'Hierarchical Common-Crawl 🍩':
            st.info("Not implemented yet...")
            
    with col2:
        st.info(f'parsing 📑...')
        parsed_info, _ = run_prediction(image)
        st.text(f'\n{information}')
        st.json(parsed_info)