app.py

import gradio as gr
import torch, os, json, requests, sys
from PIL import Image
from transformers import DonutProcessor, VisionEncoderDecoderModel, VisionEncoderDecoderConfig
from torchvision import transforms

def load_image_from_URL(url):
    res = requests.get(url)

    if res.status_code == 200:
        img = Image.open(requests.get(url, stream = True).raw)

        if img.mode == "RGBA":
            img = img.convert("RGB")

        return img

    return None

class OCRVQAModel(torch.nn.Module):
    def add_tokens(self, list_of_tokens):
        self.added_tokens.update(list_of_tokens)
        newly_added_num = self.processor.tokenizer.add_tokens(list_of_tokens)
        
        if newly_added_num > 0:
            self.donut.decoder.resize_token_embeddings(len(self.processor.tokenizer))
    
    def __init__(self, config):
        super().__init__()
        
        self.model_name_or_path = config['donut']
        self.processor_name_or_path = config['processor']
        self.config_name_or_path = config['config']
                
        self.donut_config = VisionEncoderDecoderConfig.from_pretrained(self.config_name_or_path)
        self.donut_config.encoder.image_size = [800, 600]
        self.donut_config.decoder.max_length = 64
        
        self.processor = DonutProcessor.from_pretrained(self.processor_name_or_path)
        self.donut = VisionEncoderDecoderModel.from_pretrained(self.model_name_or_path, config = self.donut_config)
        
        self.added_tokens = set([])
        self.setup()
        
    def setup(self):
        self.add_tokens(["<yes/>", "<no/>"])
        self.processor.feature_extractor.size = self.donut_config.encoder.image_size[::-1]
        self.processor.feature_extractor.do_align_long_axis = False
            
    def inference(self, image, prompt, device):
        # try:    
        self.donut.eval()
        with torch.no_grad():

            print(type(image), type(prompt), file = sys.stderr)
            image_ids = self.processor(image, return_tensors="pt").pixel_values.to(device)

            question = f'<s_docvqa><s_question>{prompt}</s_question><s_answer>'

            embedded_question = self.processor.tokenizer(
                question,
                add_special_tokens = False, 
                return_tensors = "pt"
            )["input_ids"].to(device)

            outputs = self.donut.generate(
                image_ids,
                decoder_input_ids=embedded_question,
                max_length = self.donut.decoder.config.max_position_embeddings,
                early_stopping = True,
                pad_token_id = self.processor.tokenizer.pad_token_id,
                eos_token_id = self.processor.tokenizer.eos_token_id,
                use_cache = True,
                num_beams = 1,
                bad_words_ids = [
                    [self.processor.tokenizer.unk_token_id]
                ],
                return_dict_in_generate = True
            )
            
            return self.processor.token2json(self.processor.batch_decode(outputs.sequences)[0])
        # except Exception as e:
            # raise e
            # return {
            #     'question': prompt,
            #     'answer': 'Some error occurred during inference time.'
            # }

model = OCRVQAModel({
    "donut": "ndtran/donut_ocr-vqa-200k",
    "processor": "ndtran/donut_ocr-vqa-200k",
    "config": "naver-clova-ix/donut-base-finetuned-docvqa"
})

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

def get_answer(image, url, question) -> str:
    global model, device

    if url is not None and (url.startswith('http') or url.startswith('https')):
        result = model.inference(load_image_from_URL(url), question, device)
        return result.get('answer', 'I don\'t know :<')
    
    result = model.inference(image, question, device)
    return result.get('answer', 'I don\'t know :<')


with gr.Blocks() as demo:
    with gr.Row():
        gr.Markdown(
            """
            ## Donut-OCR-VQA
            - This demo uses fine-tuned OCR-VQA-Donut model on the OCR-VQA-200k dataset to answer questions about images.
            
            ## IO description
            - Input is an image or URL that represents a book cover (recommended) and a question that asks about information on the image.
            - Output: an answer to the question.
            """
        )

    with gr.Row():
        with gr.Column():
            
            image = gr.Image(shape=(224, 224), type="pil", label="Pick an image")
            image_url = gr.Textbox(lines=1, label="Or use this option!", placeholder="Enter the image URL here")
            question = gr.Textbox(lines=5, label="Question")

            ask = gr.Button(label="Get the answer")

        with gr.Column():
            answer = gr.Label(label="Answer")

    ask.click(get_answer, inputs=[image, image_url, question], outputs=[answer])

    demo.launch()