app.py

from transformers import (
    NougatProcessor,
    VisionEncoderDecoderModel,
    TextIteratorStreamer,
)
import gradio as gr
import torch
from pathlib import Path
from pdf2image import convert_from_path
import spaces
from threading import Thread

models_supported = {
    "arabic-small-nougat": [
        NougatProcessor.from_pretrained("MohamedRashad/arabic-small-nougat"),
        VisionEncoderDecoderModel.from_pretrained("MohamedRashad/arabic-small-nougat"),
    ],
    "arabic-base-nougat": [
        NougatProcessor.from_pretrained("MohamedRashad/arabic-base-nougat"),
        VisionEncoderDecoderModel.from_pretrained(
            "MohamedRashad/arabic-base-nougat",
            torch_dtype=torch.bfloat16,
            attn_implementation={"decoder": "flash_attention_2", "encoder": "eager"},
        ),
    ],
    "arabic-large-nougat": [
        NougatProcessor.from_pretrained("MohamedRashad/arabic-large-nougat"),
        VisionEncoderDecoderModel.from_pretrained(
            "MohamedRashad/arabic-large-nougat",
            torch_dtype=torch.bfloat16,
            attn_implementation={"decoder": "flash_attention_2", "encoder": "eager"},
        ),
    ],
}


@spaces.GPU
def extract_text_from_image(image, model_name):
    print(f"Extracting text from image using model: {model_name}")
    processor, model = models_supported[model_name]
    context_length = model.decoder.config.max_position_embeddings
    torch_dtype = model.dtype
    device = "cuda" if torch.cuda.is_available() else "cpu"
    model.to(device)

    pixel_values = (
        processor(image, return_tensors="pt").pixel_values.to(torch_dtype).to(device)
    )
    streamer = TextIteratorStreamer(processor.tokenizer, skip_special_tokens=True)

    # Start generation in a separate thread
    generation_kwargs = {
        "pixel_values": pixel_values,
        "min_length": 1,
        "max_new_tokens": context_length,
        "streamer": streamer,
    }

    thread = Thread(target=model.generate, kwargs=generation_kwargs)
    thread.start()

    # Yield tokens as they become available
    output = ""
    for token in streamer:
        output += token
        yield output

    thread.join()


@spaces.GPU
def extract_text_from_pdf(pdf_path, model_name):
    processor, model = models_supported[model_name]
    context_length = model.decoder.config.max_position_embeddings
    torch_dtype = model.dtype
    device = "cuda" if torch.cuda.is_available() else "cpu"
    model.to(device)

    streamer = TextIteratorStreamer(processor.tokenizer, skip_special_tokens=True)
    print(f"Extracting text from PDF: {pdf_path}")
    images = convert_from_path(pdf_path)

    pdf_output = ""
    for image in images:
        pixel_values = (
            processor(image, return_tensors="pt")
            .pixel_values.to(torch_dtype)
            .to(device)
        )

        # Start generation in a separate thread
        generation_kwargs = {
            "pixel_values": pixel_values,
            "min_length": 1,
            "max_new_tokens": context_length,
            "streamer": streamer,
        }

        thread = Thread(target=model.generate, kwargs=generation_kwargs)
        thread.start()

        # Yield tokens as they become available
        for token in streamer:
            pdf_output += token
            yield pdf_output

        thread.join()
        pdf_output += "\n\n"
        yield pdf_output


model_description = """This is the official demo for the Arabic Nougat models. It is an end-to-end Markdown Extraction model that extracts text from images or PDFs and write them in Markdown.

There are three models available:
- [arabic-small-nougat](https://huggingface.co/MohamedRashad/arabic-small-nougat): A small model that is faster but less accurate (a finetune from [facebook/nougat-small](https://huggingface.co/facebook/nougat-small)).
- [arabic-base-nougat](https://huggingface.co/MohamedRashad/arabic-base-nougat): A base model that is more accurate but slower (a finetune from [facebook/nougat-base](https://huggingface.co/facebook/nougat-base)).
- [arabic-large-nougat](https://huggingface.co/MohamedRashad/arabic-large-nougat): The largest of the three (Made from scratch using [riotu-lab/Aranizer-PBE-86k](https://huggingface.co/riotu-lab/Aranizer-PBE-86k) tokenizer and a larger transformer decoder model).

**Disclaimer**: These models hallucinate text and are not perfect. They are trained on a mix of synthetic and real data and may not work well on all types of images.
"""

example_images = list(Path(__file__).parent.glob("*.jpeg"))

with gr.Blocks(title="Arabic Nougat") as demo:
    gr.HTML(
        "<h1 style='text-align: center'>Arabic End-to-End Structured OCR for textbooks</h1>"
    )
    gr.Markdown(model_description)

    with gr.Tab("Extract Text from Image"):
        with gr.Row():
            with gr.Column():
                input_image = gr.Image(label="Input Image", type="pil")
                model_dropdown = gr.Dropdown(
                    label="Model", choices=list(models_supported.keys()), value=None
                )
                image_submit_button = gr.Button(value="Submit", variant="primary")
            output = gr.Markdown(label="Output Markdown", rtl=True)
        image_submit_button.click(
            extract_text_from_image,
            inputs=[input_image, model_dropdown],
            outputs=output,
        )
        gr.Examples(
            example_images,
            [input_image],
            output,
            extract_text_from_image,
            cache_examples=False,
        )

    with gr.Tab("Extract Text from PDF"):
        with gr.Row():
            with gr.Column():
                pdf = gr.File(label="Input PDF", type="filepath")
                model_dropdown = gr.Dropdown(
                    label="Model", choices=list(models_supported.keys()), value=None
                )
                pdf_submit_button = gr.Button(value="Submit", variant="primary")
            output = gr.Markdown(label="Output Markdown", rtl=True)
        pdf_submit_button.click(
            extract_text_from_pdf, inputs=[pdf, model_dropdown], outputs=output
        )

demo.queue().launch(share=False)