app.py

import gradio as gr
from transformers import AutoModelForQuestionAnswering, AutoTokenizer, LayoutLMv3ImageProcessor

model_name = "TusharGoel/LiLT-Document-QA"
tokenizer = AutoTokenizer.from_pretrained(model_name, apply_ocr = True)
image_processor = LayoutLMv3ImageProcessor()

model = AutoModelForQuestionAnswering.from_pretrained(model_name)
model.eval()

def qna(image, question):
    
    res = image_processor(image, apply_ocr = True)
    words = res["words"][0]
    boxes = res["boxes"][0]

    encoding = tokenizer(question, words, boxes = boxes, return_token_type_ids=True, return_tensors="pt", truncation=True, padding="max_length")

    word_ids = encoding.word_ids(0)
    outputs = model(**encoding)

    start_scores = outputs.start_logits
    end_scores = outputs.end_logits

    start, end = word_ids[start_scores.argmax(-1).item()], word_ids[end_scores.argmax(-1).item()]

    answer = " ".join(words[start : end + 1])
    
    
    return answer


img = gr.Image(source="upload", label="Image")
question = gr.Text(label="Question")
label = gr.Label(label="label")

iface = gr.Interface(fn=qna, inputs=[img, question], outputs=label, title="LiLT - Document Question Answering", allow_duplication=True)
iface.launch()