app.py

import numpy as np
import gradio as gr
from transformers import AutoFeatureExtractor, AutoTokenizer, VisionEncoderDecoderModel
import re
import jaconv

#load model
model_path = "model/"
feature_extractor = AutoFeatureExtractor.from_pretrained(model_path)
tokenizer = AutoTokenizer.from_pretrained(model_path)
model = VisionEncoderDecoderModel.from_pretrained(model_path)

examples = ['examples/01.png', 'examples/02.png', 'examples/03.png',
            'examples/04.png', 'examples/05.png', 'examples/06.png',
            'examples/07.png'
            ]
            
def post_process(text):
    text = ''.join(text.split())
    text = text.replace('…', '...')
    text = re.sub('[・.]{2,}', lambda x: (x.end() - x.start()) * '.', text)
    text = jaconv.h2z(text, ascii=True, digit=True)
    return text

def infer(image):
    image = image.convert('L').convert('RGB')
    pixel_values = feature_extractor(image, return_tensors="pt").pixel_values
    ouput = model.generate(pixel_values)[0]
    text = tokenizer.decode(ouput, skip_special_tokens=True)
    text = post_process(text)
    return text


iface = gr.Interface(
    fn=infer,
    inputs=[gr.inputs.Image(label="Input", type="pil")],
    outputs="text",
    layout="horizontal",
    theme="huggingface",
    title="Optical Character Recognition for Japanese Text",
    description="A simple interface for OCR from Japanese manga",
    article= "Author: <a href=\"https://huggingface.co/vumichien\">Vu Minh Chien</a>. ",
    allow_flagging='never',
    examples=examples,
    cache_examples=True,
)
iface.launch(enable_queue=True)