ocr_full_pdf12.py

from transformers import AutoModel, AutoTokenizer
import torch
import os
from PIL import Image
import time

# Force CPU for stability on Mac
device = "cpu"
print(f"Using device: {device}")

# Patch to avoid CUDA calls in custom code
torch.Tensor.cuda = lambda self, *args, **kwargs: self.to(device)
torch.nn.Module.cuda = lambda self, *args, **kwargs: self.to(device)

model_name = 'deepseek-ai/DeepSeek-OCR-2'

def ocr_full_document():
    print(f"Loading tokenizer...")
    tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True)

    print(f"Loading model...")
    model = AutoModel.from_pretrained(
        model_name, 
        trust_remote_code=True, 
        use_safetensors=True
    )
    model = model.eval()
    
    # Overwrite bfloat16 to float32 for CPU compatibility
    torch.bfloat16 = torch.float32

    image_dir = "doc_images_full"
    output_dir = "ocr_results_pdf12"
    os.makedirs(output_dir, exist_ok=True)

    # Get images sorted by page number
    import re
    def get_page_num(filename):
        match = re.search(r'page_(\d+)', filename)
        return int(match.group(1)) if match else 0

    images = sorted([f for f in os.listdir(image_dir) if f.endswith(".png")], key=get_page_num)
    
    full_markdown = []
    
    for i, img_name in enumerate(images):
        img_path = os.path.join(image_dir, img_name)
        print(f"\n[{i+1}/{len(images)}] Processing page {get_page_num(img_name)}...")
        
        prompt = "<image>\nFree OCR. "
        
        start_time = time.time()
        try:
            with torch.no_grad():
                res = model.infer(
                    tokenizer, 
                    prompt=prompt, 
                    image_file=img_path, 
                    output_path=output_dir,
                    base_size=1024, 
                    image_size=768,
                    crop_mode=False,
                    eval_mode=True
                )
            
            elapsed = time.time() - start_time
            print(f"  Done in {elapsed:.2f}s")
            
            # Save individual page result
            page_file = os.path.join(output_dir, f"{img_name}.md")
            with open(page_file, "w") as f:
                f.write(res)
            
            full_markdown.append(f"## Page {get_page_num(img_name)}\n\n{res}\n\n---\n")
            
        except Exception as e:
            print(f"  Failed: {e}")
            full_markdown.append(f"## Page {get_page_num(img_name)}\n\n[OCR FAILED]\n\n---\n")

    # Save combined result
    combined_file = os.path.join(output_dir, "full_document.md")
    with open(combined_file, "w") as f:
        f.write("# OCR Result for pdf12_un.pdf\n\n")
        f.write("".join(full_markdown))
    
    print(f"\nCompleted! Full result saved to: {combined_file}")

if __name__ == "__main__":
    ocr_full_document()