app.py

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

processor = DonutProcessor.from_pretrained("naver-clova-ix/donut-base-finetuned-rvlcdip")
model = VisionEncoderDecoderModel.from_pretrained("naver-clova-ix/donut-base-finetuned-rvlcdip")

def ClassificateDocs(pathimage):
  image = Image.open(pathimage)
  pixel_values = processor(image, return_tensors="pt").pixel_values
  task_prompt = "<s_rvlcdip>"
  decoder_input_ids = processor.tokenizer(task_prompt, add_special_tokens=False, return_tensors="pt").input_ids
  device = "cuda" if torch.cuda.is_available() else "cpu"
  model.to(device)
  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)

processor_prs= DonutProcessor.from_pretrained("naver-clova-ix/donut-base-finetuned-cord-v2")
model_prs = VisionEncoderDecoderModel.from_pretrained("naver-clova-ix/donut-base-finetuned-cord-v2")

def ProcessBill(pathimage ):
  image = Image.open(pathimage)
  pixel_values = processor_prs(image, return_tensors="pt").pixel_values
  task_prompt = "<s_cord-v2>"
  decoder_input_ids = processor_prs.tokenizer(task_prompt, add_special_tokens=False, return_tensors="pt")["input_ids"]
  device = "cuda" if torch.cuda.is_available() else "cpu"
  model_prs.to(device)
  outputs = model_prs.generate(pixel_values.to(device),
                               decoder_input_ids=decoder_input_ids.to(device),
                               max_length=model_prs.decoder.config.max_position_embeddings,
                               early_stopping=True,
                               pad_token_id=processor_prs.tokenizer.pad_token_id,
                               eos_token_id=processor_prs.tokenizer.eos_token_id,
                               use_cache=True,
                               num_beams=1,
                               bad_words_ids=[[processor_prs.tokenizer.unk_token_id]],
                               return_dict_in_generate=True,
                               output_scores=True,)
  sequence = processor_prs.batch_decode(outputs.sequences)[0]
  sequence = sequence.replace(processor_prs.tokenizer.eos_token, "").replace(processor_prs.tokenizer.pad_token, "")
  sequence = re.sub(r"<.*?>", "", sequence, count=1).strip()  # remove first task start token
  return processor_prs.token2json(sequence)

processor_qa= DonutProcessor.from_pretrained("naver-clova-ix/donut-base-finetuned-docvqa")
model_qa = VisionEncoderDecoderModel.from_pretrained("naver-clova-ix/donut-base-finetuned-docvqa")
def QAsBill(pathimage,question="When is the coffee break?" ):
  image = Image.open(pathimage)
  pixel_values = processor_qa(image, return_tensors="pt").pixel_values
  task_prompt = "<s_docvqa><s_question>{user_input}</s_question><s_answer>"
  prompt = task_prompt.replace("{user_input}", question)
  decoder_input_ids = processor_qa.tokenizer(prompt, add_special_tokens=False, return_tensors="pt").input_ids

  device = "cuda" if torch.cuda.is_available() else "cpu"
  model_qa.to(device)
  outputs = model_qa.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_qa.tokenizer.pad_token_id,
    eos_token_id=processor_qa.tokenizer.eos_token_id,
    use_cache=True,
    bad_words_ids=[[processor_qa.tokenizer.unk_token_id]],
    return_dict_in_generate=True,
  )

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


demo = gr.Blocks()

gradio_app_cls = gr.Interface(
    fn=ClassificateDocs,
    inputs=[
        gr.Image(type='filepath')
    ],
    outputs="text",
)
gradio_app_prs = gr.Interface(
    fn=ProcessBill,
    inputs=[
        gr.Image(type='filepath')
    ],
    outputs="text",
)
gradio_app_qa = gr.Interface(
    fn=QAsBill,
    inputs=[
        gr.Image(type='filepath'),
        gr.Text()
    ],
    outputs="text",
)


demo = gr.TabbedInterface([gradio_app_cls, gradio_app_prs,gradio_app_qa], ["class", "parse","QA"])

if __name__ == "__main__":
    demo.launch()