app.py

# Importing necessary packages
import torch # PyTorch used for executing deep learning functions
from PIL import Image, ImageTk # to display the image from the encoded pixels
import gradio as gr
from transformers import TrOCRProcessor, VisionEncoderDecoderModel # importing the TrOCR processor representing the visual feature extrcator and tokenizer of the TrOCR model, and the TrOCR model
import os

processor = TrOCRProcessor.from_pretrained("microsoft/trocr-base-handwritten")
auth_token = os.environ.get("TOKEN_FROM_SECRET") or True
model = VisionEncoderDecoderModel.from_pretrained("sk2003/hist-trocr", use_auth_token=auth_token)
# def process_image(image):
#     # prepare image
#     pixel_values = processor(image, return_tensors="pt").pixel_values

#     # generate
#     generated_ids = model.generate(pixel_values)

#     # decode
#     generated_text = processor.batch_decode(generated_ids, skip_special_tokens=True)[0]

#     return generated_text
def inference_on_image(image):
    pixel_values = processor(image, return_tensors="pt").pixel_values
    pred = model.generate(pixel_values, max_new_tokens=100)
    dec_pred = processor.batch_decode(pred, skip_special_tokens=True)[0]
    return dec_pred

title = "Hist-TrOCR"
description = "Interactive demo of Hist-TrOCR, a fine-tuned version of Microsoft's TrOCR which is an end-to-end transformer model used for recognition of text from single-line or word images. It has been fine-tuned on historical text images. Upload an image (or select from the given samples) and click 'Submit' to get the transcription. Results may take a few seconds to show up."

iface = gr.Interface(fn=inference_on_image, 
                     inputs=gr.inputs.Image(type="pil"), 
                     outputs=gr.outputs.Textbox(),
                     title=title,
                     examples=[["309-35.png"],["270-01-03.png"],["v211285.b750.s69.jpg"]],
                     description=description)
iface.launch(debug=True)