app.py

from PyPDF2 import PdfReader
import gradio as gr
from docling.document_converter import DocumentConverter, PdfFormatOption
from docling.datamodel.pipeline_options import PdfPipelineOptions
from docling.datamodel.base_models import InputFormat
from paddleocr import PPStructureV3
from pdf2image import convert_from_path
import numpy as np
import torch
from docling_core.types.doc import DoclingDocument
from docling_core.types.doc.document import DocTagsDocument
from transformers import AutoProcessor, AutoModelForVision2Seq
from transformers.image_utils import load_image
from pathlib import Path
import time
import os

DEVICE = "cuda" if torch.cuda.is_available() else "cpu"

# Optimize for CPU: set float32 and use all CPU threads
if DEVICE == "cpu":
    torch.set_num_threads(os.cpu_count() or 1)
    smoldocling_dtype = torch.float32
else:
    smoldocling_dtype = torch.bfloat16

pipeline_options = PdfPipelineOptions(enable_remote_services=True)
converter = DocumentConverter(
    format_options={
        InputFormat.PDF: PdfFormatOption(pipeline_options=pipeline_options)
    }
)

pipeline = PPStructureV3()

processor = AutoProcessor.from_pretrained("ds4sd/SmolDocling-256M-preview")
model = AutoModelForVision2Seq.from_pretrained(
    "ds4sd/SmolDocling-256M-preview",
    torch_dtype=smoldocling_dtype,
    _attn_implementation="flash_attention_2" if DEVICE == "cuda" else "eager",
).to(DEVICE)

def get_pdf_page_count(pdf_path):
    reader = PdfReader(pdf_path)
    return len(reader.pages)

def get_page_image(pdf_path, page_num):
    start = time.time()
    images = convert_from_path(pdf_path, first_page=page_num, last_page=page_num)
    page_image = images[0]
    runtime = time.time() - start
    return page_image, f"{runtime:.2f} s"

def get_docling_ocr(pdf_path, page_num):
    start = time.time()
    result = converter.convert(pdf_path, page_range=(page_num, page_num))
    markdown_text_docling = result.document.export_to_markdown()
    runtime = time.time() - start
    return markdown_text_docling, f"{runtime:.2f} s"

def get_paddle_ocr(pdf_path, page_num):
    start = time.time()
    page_image = get_page_image(pdf_path, page_num)[0]
    output = pipeline.predict(input=np.array(page_image))
    markdown_list = []
    for res in output:
        md_info = res.markdown
        markdown_list.append(md_info)
    markdown_text_paddleOCR = pipeline.concatenate_markdown_pages(markdown_list)
    runtime = time.time() - start
    return markdown_text_paddleOCR, f"{runtime:.2f} s"

def get_smoldocling_ocr(pdf_path, page_num):
    start = time.time()
    page_image = get_page_image(pdf_path, page_num)[0]
    image = load_image(page_image)
    messages = [
        {
            "role": "user",
            "content": [
                {"type": "image"},
                {"type": "text", "text": "Convert this page to docling."}
            ]
        },
    ]
    prompt = processor.apply_chat_template(messages, add_generation_prompt=True)
    inputs = processor(text=prompt, images=[image], return_tensors="pt")
    inputs = inputs.to(DEVICE)
    with torch.no_grad():
        generated_ids = model.generate(**inputs, max_new_tokens=1500, do_sample=False, num_beams=1, temperature=1.0)
    prompt_length = inputs.input_ids.shape[1]
    trimmed_generated_ids = generated_ids[:, prompt_length:]
    doctags = processor.batch_decode(
        trimmed_generated_ids,
        skip_special_tokens=False,
    )[0].lstrip()
    doctags_doc = DocTagsDocument.from_doctags_and_image_pairs([doctags], [image])
    doc = DoclingDocument.load_from_doctags(doctags_doc, document_name="Document")
    markdown_text_smoldocling = doc.export_to_markdown()
    runtime = time.time() - start
    return markdown_text_smoldocling, f"{runtime:.2f} s"

title = "OCR Arena"
description = "A simple Gradio interface to extract text from PDFs and compare OCR models"
examples = [["data/amazon-10-k-2024.pdf"],
            ["data/goog-10-k-2023.pdf"]]

with gr.Blocks(theme=gr.themes.Glass()) as demo:
    gr.Markdown(f"# {title}\n{description}")
    with gr.Column():
        pdf = gr.File(label="Input PDFs", file_types=[".pdf"])

        @gr.render(inputs=pdf)
        def show_slider(pdf_path):
            if pdf_path is None:
                page_num = gr.Markdown("## No Input Provided")
            else:
                page_count = get_pdf_page_count(pdf_path)
                page_num = gr.Slider(1, page_count, value=1, step=1, label="Page Number")

                with gr.Row():
                    clear_btn = gr.ClearButton(components=[pdf, page_num])
                    submit_btn = gr.Button("Submit", variant='primary')

                submit_btn.click(get_page_image, inputs=[pdf, page_num], outputs=[original, original_runtime]).then(    
                    get_docling_ocr, inputs=[pdf, page_num], outputs=[docling_ocr_out, docling_ocr_runtime]).then(
                    get_paddle_ocr, inputs=[pdf, page_num], outputs=[paddle_ocr_out, paddle_ocr_runtime]).then(
                    get_smoldocling_ocr, inputs=[pdf, page_num], outputs=[smoldocling_ocr_out, smoldocling_ocr_runtime])

    with gr.Column():
        with gr.Row():
            with gr.Column():
                original = gr.Image(width=640, height=640, label="Original Page", interactive=False)
                original_runtime = gr.Textbox(label="Image Extraction Time", type="text", interactive=False)
            with gr.Column():
                docling_ocr_out = gr.Textbox(label="Docling OCR Output", type="text", interactive=False)
                docling_ocr_runtime = gr.Textbox(label="Docling OCR Time", type="text", interactive=False)
        with gr.Row():
            with gr.Column():
                paddle_ocr_out = gr.Textbox(label="Paddle OCR Output", type="text", interactive=False)
                paddle_ocr_runtime = gr.Textbox(label="Paddle OCR Time", type="text", interactive=False)
            with gr.Column():
                smoldocling_ocr_out = gr.Textbox(label="SmolDocling OCR Output", type="text", interactive=False)
                smoldocling_ocr_runtime = gr.Textbox(label="SmolDocling OCR Time", type="text", interactive=False)

    examples_obj = gr.Examples(examples=examples, inputs=[pdf])

demo.launch()