app.py

import re

from transformers import DonutProcessor, VisionEncoderDecoderModel
import gradio as gr
import torch
from PIL import Image

processor = DonutProcessor.from_pretrained("ewfian/donut_cn_invoice")
model = VisionEncoderDecoderModel.from_pretrained("ewfian/donut_cn_invoice")

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

task_prompt = "<s_totalAmountInWords>"
decoder_input_ids = processor.tokenizer(task_prompt, add_special_tokens=False, return_tensors="pt").input_ids

def process_document(image):

    pixel_values = processor(image, return_tensors="pt").pixel_values

    outputs = model.generate(
        pixel_values.to(device),
        decoder_input_ids=decoder_input_ids.to(device),
        max_length=model.decoder.config.max_position_embeddings,
        pad_token_id=processor.tokenizer.pad_token_id,
        eos_token_id=processor.tokenizer.eos_token_id,
        use_cache=True,
        bad_words_ids=[[processor.tokenizer.unk_token_id]],
        return_dict_in_generate=True,
    )

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

demo = gr.Interface(
    fn=process_document,
    inputs="image",
    outputs="json",
    title="Demo: Donut 🍩 for Chinese Invioce Parsing",
    cache_examples=False)

demo.launch(server_name="0.0.0.0")