app.py

# import re
# import gradio as gr
# from transformers import AutoProcessor, AutoModelForImageTextToText
# from PIL import Image

# # Load model & processor once at startup
# processor = AutoProcessor.from_pretrained("ds4sd/SmolDocling-256M-preview")
# model = AutoModelForImageTextToText.from_pretrained("ds4sd/SmolDocling-256M-preview")

# def smoldocling_readimage(image, prompt_text="Convert to docling"):
#     messages = [
#         {"role": "user", "content": [{"type": "image"}, {"type": "text", "text": prompt_text}]}
#     ]
#     prompt = processor.apply_chat_template(messages, add_generation_prompt=True)
#     inputs = processor(text=prompt, images=[image], return_tensors="pt")
#     outputs = model.generate(**inputs, max_new_tokens=1024)
#     prompt_length = inputs.input_ids.shape[1]
#     generated = outputs[:, prompt_length:]
#     result = processor.batch_decode(generated, skip_special_tokens=False)[0]
#     return result.replace("<end_of_utterance>", "").strip()

# def extract_numbers(docling_text):
#     # Extract all floating numbers from the docling text using regex
#     numbers = re.findall(r"[-+]?\d*\.\d+|\d+", docling_text)
#     return list(map(float, numbers))

# def compare_outputs(img1, img2):
#     # Extract docling text from both images
#     output1 = smoldocling_readimage(img1)
#     output2 = smoldocling_readimage(img2)
    
#     # Extract numbers from both outputs
#     nums1 = extract_numbers(output1)
#     nums2 = extract_numbers(output2)
    
#     # Compare numbers — find matching count based on position
#     length = min(len(nums1), len(nums2))
#     matches = sum(1 for i in range(length) if abs(nums1[i] - nums2[i]) < 1e-3)
    
#     # Calculate similarity accuracy percentage
#     total = max(len(nums1), len(nums2))
#     accuracy = (matches / total) * 100 if total > 0 else 0
    
#     # Prepare result text
#     result_text = (
#         f"Output for Image 1:\n{output1}\n\n"
#         f"Output for Image 2:\n{output2}\n\n"
#         f"Similarity Accuracy: {accuracy:.2f}%\n"
#         f"Matching Values: {matches} out of {total}"
#     )
#     return result_text

# # Gradio UI: take 2 images, output similarity report
# demo = gr.Interface(
#     fn=compare_outputs,
#     inputs=[
#         gr.Image(type="pil", label="Upload Image 1"),
#         gr.Image(type="pil", label="Upload Image 2"),
#     ],
#     outputs="text",
#     title="SmolDocling Image Comparison",
#     description="Upload two document images. This app extracts data from both and compares similarity."
# )

# demo.launch()


import re
import gradio as gr
from transformers import AutoProcessor, AutoModelForImageTextToText
from PIL import Image

# Load model & processor once at startup
processor = AutoProcessor.from_pretrained("ds4sd/SmolDocling-256M-preview")
model = AutoModelForImageTextToText.from_pretrained("ds4sd/SmolDocling-256M-preview")

def smoldocling_readimage(image, prompt_text="Convert to docling"):
    messages = [
        {"role": "user", "content": [{"type": "image"}, {"type": "text", "text": prompt_text}]}
    ]
    prompt = processor.apply_chat_template(messages, add_generation_prompt=True)
    inputs = processor(text=prompt, images=[image], return_tensors="pt")
    outputs = model.generate(**inputs, max_new_tokens=1024)
    prompt_length = inputs.input_ids.shape[1]
    generated = outputs[:, prompt_length:]
    result = processor.batch_decode(generated, skip_special_tokens=False)[0]
    return result.replace("<end_of_utterance>", "").strip()

def extract_numbers(docling_text):
    # Extract all floating numbers from the docling text
    numbers = re.findall(r"[-+]?\d*\.\d+|\d+", docling_text)
    return list(map(float, numbers))

def compare_outputs(img1, img2):
    # Get outputs
    output1 = smoldocling_readimage(img1)
    output2 = smoldocling_readimage(img2)
    
    # Extract numbers
    nums1 = extract_numbers(output1)
    nums2 = extract_numbers(output2)
    
    length = min(len(nums1), len(nums2))
    matches = 0
    mismatches = []
    
    for i in range(length):
        if abs(nums1[i] - nums2[i]) < 1e-3:
            matches += 1
        else:
            mismatches.append(f"Pos {i+1}: {nums1[i]}  ≠  {nums2[i]}")
    
    total = max(len(nums1), len(nums2))
    accuracy = (matches / total) * 100 if total > 0 else 0
    
    mismatch_text = "\n".join(mismatches) if mismatches else "✅ All values match."
    
    result_text = (
        f"📄 Output for Image 1:\n{output1}\n\n"
        f"📄 Output for Image 2:\n{output2}\n\n"
        f"🔍 Similarity Accuracy: {accuracy:.2f}%\n"
        f"✅ Matching Values: {matches} / {total}\n"
        f"❌ Mismatches:\n{mismatch_text}"
    )
    return result_text

# Gradio UI
demo = gr.Interface(
    fn=compare_outputs,
    inputs=[
        gr.Image(type="pil", label="Upload Image 1"),
        gr.Image(type="pil", label="Upload Image 2"),
    ],
    outputs="text",
    title="SmolDocling Image Comparison",
    description="Upload two document images to extract values and compare similarity, with detailed mismatches."
)

demo.launch()