import os import hashlib import spaces import re import time import click import gradio as gr from io import BytesIO from PIL import Image from loguru import logger from pathlib import Path import torch from transformers import AutoProcessor, Qwen2_5_VLForConditionalGeneration from transformers.image_utils import load_image import fitz import html2text import markdown import tempfile from typing import Optional, Tuple # --- Constants and Setup --- pdf_suffixes = [".pdf"] image_suffixes = [".png", ".jpeg", ".jpg"] device = "cuda" if torch.cuda.is_available() else "cpu" # --- Model and Processor Initialization --- logger.info(f"Using device: {device}") # Model 1: Logics-Parsing MODEL_ID_1 = "Logics-MLLM/Logics-Parsing" logger.info(f"Loading model 1: {MODEL_ID_1}") processor_1 = AutoProcessor.from_pretrained(MODEL_ID_1, trust_remote_code=True) model_1 = Qwen2_5_VLForConditionalGeneration.from_pretrained( MODEL_ID_1, trust_remote_code=True, torch_dtype=torch.float16 if device == "cuda" else torch.float32 ).to(device).eval() logger.info(f"Model '{MODEL_ID_1}' loaded successfully.") # Model 2: Gliese-OCR-7B-Post1.0 MODEL_ID_2 = "prithivMLmods/Gliese-OCR-7B-Post1.0" logger.info(f"Loading model 2: {MODEL_ID_2}") processor_2 = AutoProcessor.from_pretrained(MODEL_ID_2, trust_remote_code=True) model_2 = Qwen2_5_VLForConditionalGeneration.from_pretrained( MODEL_ID_2, trust_remote_code=True, torch_dtype=torch.float16 if device == "cuda" else torch.float32 ).to(device).eval() logger.info(f"Model '{MODEL_ID_2}' loaded successfully.") @spaces.GPU def parse_page(image: Image.Image, model_name: str) -> str: """ Parses a single document page image using the selected model. """ # Select the appropriate model and processor based on the choice if model_name == "Logics-Parsing": current_processor = processor_1 current_model = model_1 elif model_name == "Gliese-OCR-7B-Post1.0": current_processor = processor_2 current_model = model_2 else: raise ValueError(f"Unknown model choice: {model_name}") messages = [ { "role": "user", "content": [ {"type": "image", "image": image}, {"type": "text", "text": "Parse this document page into a clean, structured HTML representation. Preserve the logical structure with appropriate tags for content blocks such as paragraphs (

), headings (

-

), tables (), figures (
), formulas (), and others. Include category tags, and filter out irrelevant elements like headers and footers."}, ], }, ] prompt_full = current_processor.apply_chat_template(messages, tokenize=False, add_generation_prompt=True) inputs = current_processor( text=[prompt_full], images=[image], return_tensors="pt", padding=True ).to(device) with torch.no_grad(): generated_ids = current_model.generate( **inputs, max_new_tokens=2048, temperature=0.1, top_p=0.9, do_sample=True, repetition_penalty=1.05 ) generated_ids_trimmed = [ out_ids[len(in_ids):] for in_ids, out_ids in zip(inputs.input_ids, generated_ids) ] output_text = current_processor.batch_decode( generated_ids_trimmed, skip_special_tokens=True, clean_up_tokenization_spaces=False )[0] return output_text def convert_pdf_to_images_fitz(pdf_path: str, dpi: int = 200) -> list: """ Converts a PDF file to a list of PIL Images using PyMuPDF (fitz). """ images = [] try: pdf_document = fitz.open(pdf_path) zoom = dpi / 72.0 mat = fitz.Matrix(zoom, zoom) for page_num in range(len(pdf_document)): page = pdf_document.load_page(page_num) pix = page.get_pixmap(matrix=mat) img_data = pix.tobytes("png") image = Image.open(BytesIO(img_data)) images.append(image) pdf_document.close() except Exception as e: logger.error(f"Failed to convert PDF using PyMuPDF: {e}") raise return images async def pdf_parse(file_path: str, model_choice: str): """ Main parsing function that orchestrates the PDF processing pipeline. """ if not file_path: logger.warning("File path is None.") return "

Please upload a file first.

", "", "", None, "Error: No file provided", None, "No file loaded" logger.info(f'Processing file: {file_path} with model: {model_choice}') start_time = time.time() try: pages = convert_pdf_to_images_fitz(file_path, dpi=200) if not pages: raise ValueError("Could not extract any pages from the PDF.") html_parts = [] for i, page in enumerate(pages): logger.info(f"Parsing page {i+1}/{len(pages)}") # Pass the model choice to the parsing function html = parse_page(page, model_choice) html_parts.append(f'\n{html}') full_html = '\n'.join(html_parts) parsing_time = time.time() - start_time mmd = html2text.html2text(full_html) mmd_html = markdown.markdown(mmd, extensions=['fenced_code', 'tables']) with tempfile.NamedTemporaryFile(mode='w', suffix='.md', delete=False, encoding='utf-8') as f: f.write(mmd) md_path = f.name cost_time_str = f'Total processing time: {parsing_time:.2f}s' preview_image = pages[0] page_info_html = f'
Page 1 / {len(pages)}
' return mmd_html, mmd, full_html, md_path, cost_time_str, preview_image, page_info_html except Exception as e: logger.error(f"Parsing failed: {e}", exc_info=True) error_html = f"

An error occurred during processing:

{str(e)}

" return error_html, "", "", None, f"Error: {str(e)}", None, "Error processing" def show_pdf_preview_as_image(file_path: Optional[str]) -> Tuple[Optional[Image.Image], str]: """ Generates a PIL Image preview of the first page of a PDF or image file and provides page count information. """ if not file_path: return None, '
No file loaded
' page_info_html = '
Page 1 / 1
' try: if Path(file_path).suffix.lower() in image_suffixes: return Image.open(file_path).convert("RGB"), page_info_html elif Path(file_path).suffix.lower() == '.pdf': doc = fitz.open(file_path) page_count = len(doc) page_info_html = f'
Page 1 / {page_count}
' if page_count > 0: page = doc.load_page(0) zoom = 200 / 72.0 mat = fitz.Matrix(zoom, zoom) pix = page.get_pixmap(matrix=mat) img = Image.open(BytesIO(pix.tobytes("png"))) doc.close() return img, page_info_html doc.close() except Exception as e: logger.error(f"Failed to create file preview: {e}") return None, '
Failed to load preview
' def clear_all(): """Clears all input and output components in the UI.""" return ( None, None, "

Results will be displayed here after processing.

", "", "", None, "", '
No file loaded
' ) @click.command() def main(): """ Sets up and launches the Gradio user interface for the Logics-Parsing app. """ css = """ .main-container { max-width: 1400px; margin: 0 auto; } .header-text { text-align: center; color: #2c3e50; margin-bottom: 20px; } .process-button { border: none !important; color: white !important; font-weight: bold !important; background-color: blue !important;} .process-button:hover { background-color: darkblue !important; transform: translateY(-2px) !important; box-shadow: 0 4px 8px rgba(0,0,0,0.2) !important; } .page-info { text-align: center; padding: 8px 16px; border-radius: 20px; font-weight: bold; margin: 10px 0; } """ with gr.Blocks(theme="bethecloud/storj_theme", css=css, title="Logics-Parsing Demo") as demo: # Header gr.HTML("""

📄 Logics-Parsing: Structured Document Analysis

An advanced Vision Language Model to parse documents and images into clean HTML and Markdown.

🤗 Model Page 💻 GitHub 📝 Arxiv Paper
""") with gr.Row(elem_classes=["main-container"]): # Left column for inputs and controls with gr.Column(scale=1): model_choice = gr.Dropdown( choices=["Logics-Parsing", "Gliese-OCR-7B-Post1.0"], label="Select Model⚡️", value="Logics-Parsing" ) file_input = gr.File( label="Upload PDF or Image", file_types=[".pdf", ".jpg", ".jpeg", ".png"], type="filepath" ) image_preview = gr.Image( label="Preview", type="pil", interactive=False, height=280 ) with gr.Row(): prev_page_btn = gr.Button("◀ Previous", size="md") page_info = gr.HTML('
No file loaded
') next_page_btn = gr.Button("Next ▶", size="md") example_root = "examples" if os.path.exists(example_root) and os.path.isdir(example_root): example_files = [ os.path.join(example_root, f) for f in os.listdir(example_root) if f.endswith(tuple(pdf_suffixes + image_suffixes)) ] if example_files: with gr.Accordion("Open Examples⚙️", open=False): gr.Examples( examples=example_files, inputs=file_input, examples_per_page=10, ) with gr.Accordion("Other Details🕧", open=False): output_file = gr.File(label='Download Markdown Result', interactive=False) cost_time = gr.Text(label='Time Cost', interactive=False) process_btn = gr.Button( "🚀 Process Document", variant="primary", elem_classes=["process-button"], size="lg" ) clear_btn = gr.Button("🗑️ Clear All", variant="secondary") # Right column for results with gr.Column(scale=2): with gr.Tabs(): with gr.Tab("Markdown Source"): mmd = gr.TextArea(lines=27, show_copy_button=True, label="Markdown Source", interactive=True) with gr.Tab("Markdown Rendering"): mmd_html = gr.TextArea( lines=27, label='Markdown Rendering', show_copy_button=True ) with gr.Tab("Generated HTML"): raw_html = gr.TextArea(lines=27, show_copy_button=True, label="Generated HTML") # --- Event Handlers --- file_input.change( fn=show_pdf_preview_as_image, inputs=[file_input], outputs=[image_preview, page_info], show_progress="full" ) process_btn.click( fn=pdf_parse, inputs=[file_input, model_choice], outputs=[mmd_html, mmd, raw_html, output_file, cost_time, image_preview, page_info], concurrency_limit=15, show_progress="full" ) clear_btn.click( fn=clear_all, outputs=[ file_input, image_preview, mmd_html, mmd, raw_html, output_file, cost_time, page_info ] ) demo.queue().launch(debug=True, show_error=True) if __name__ == '__main__': if not os.path.exists("examples"): os.makedirs("examples") logger.info("Created 'examples' directory. Please add some sample PDF/image files there.") main()