app.py

import gradio as gr
import cv2
import numpy as np
import onnxruntime as ort
from huggingface_hub import hf_hub_download

# --- STEP 1: Download the ONNX Model ---
print("Downloading ONNX model...")
model_path = hf_hub_download(repo_id="alex-dinh/PP-DocLayoutV3-ONNX", filename="model.onnx")
print(f"Model downloaded to: {model_path}")

# --- STEP 2: Initialize ONNX Engine ---
# This loads the AI "brain" directly without needing Paddle
session = ort.InferenceSession(model_path)
input_names = [i.name for i in session.get_inputs()]
output_names = [o.name for o in session.get_outputs()]

# Define labels map (Standard for PP-DocLayout)
LABELS = {1: "Text", 2: "Title", 3: "List", 4: "Table", 5: "Figure"}

def preprocess_image(image, target_size=(800, 800)):
    """
    Prepares the image exactly how the AI expects it (Resize -> Normalize).
    """
    h, w = image.shape[:2]
    
    # 1. Resize
    # We do NOT keep aspect ratio for the input blob, but we keep scales to fix boxes later
    img_resized = cv2.resize(image, target_size)
    
    # 2. Normalize (Standard ImageNet mean/std)
    img_data = img_resized.astype(np.float32) / 255.0
    mean = np.array([0.485, 0.456, 0.406], dtype=np.float32)
    std = np.array([0.229, 0.224, 0.225], dtype=np.float32)
    img_data = (img_data - mean) / std
    
    # 3. Transpose to (Batch, Channel, Height, Width)
    img_data = img_data.transpose(2, 0, 1)[None, :, :, :]
    
    # Calculate scale factors to map detections back to original image
    scale_factor = np.array([target_size[0] / h, target_size[1] / w], dtype=np.float32).reshape(1, 2)
    
    return img_data, scale_factor

def analyze_layout(input_image):
    if input_image is None:
        return None, "No image uploaded"

    # Convert PIL to Numpy/OpenCV
    image_np = np.array(input_image)
    orig_h, orig_w = image_np.shape[:2]

    # --- INFERENCE ---
    input_blob, scale_factor = preprocess_image(image_np)
    
    # ONNX Runtime inputs
    inputs = {
        input_names[0]: input_blob,           # The image data
        input_names[1]: scale_factor          # The resize scale
    }
    
    # Run!
    outputs = session.run(output_names, inputs)
    
    # --- POST-PROCESSING ---
    # Output format is typically [Batch, N, 6] -> [Class, Score, X1, Y1, X2, Y2]
    detections = outputs[0]  
    
    viz_image = image_np.copy()
    log = []

    for det in detections:
        class_id = int(det[0])
        score = det[1]
        bbox = det[2:]

        if score < 0.5: continue # Filter weak detections

        # Map labels
        label_name = LABELS.get(class_id, "Unknown")
        
        # Coordinates
        x1, y1, x2, y2 = map(int, bbox)
        
        # Color coding
        color = (0, 255, 0) # Green
        if label_name == "Title": color = (0, 0, 255)
        elif label_name == "Table": color = (255, 255, 0)
        elif label_name == "Figure": color = (255, 0, 0)

        # Draw
        cv2.rectangle(viz_image, (x1, y1), (x2, y2), color, 3)
        cv2.putText(viz_image, f"{label_name} {score:.2f}", (x1, y1-10), 
                   cv2.FONT_HERSHEY_SIMPLEX, 0.6, color, 2)
        
        log.append(f"Found {label_name} at [{x1}, {y1}, {x2}, {y2}] (Conf: {score:.2f})")

    return viz_image, "\n".join(log)

with gr.Blocks(title="ONNX Layout Analysis") as demo:
    gr.Markdown("## ⚡ Fast V3 Layout Analysis (ONNX)")
    gr.Markdown("Uses **PP-DocLayoutV3** via ONNX Runtime. No Paddle dependencies.")
    
    with gr.Row():
        with gr.Column():
            input_img = gr.Image(type="pil", label="Input Document")
            submit_btn = gr.Button("Analyze Layout", variant="primary")
        
        with gr.Column():
            output_img = gr.Image(label="Layout Visualization")
            output_log = gr.Textbox(label="Detections", lines=10)

    submit_btn.click(fn=analyze_layout, inputs=input_img, outputs=[output_img, output_log])

if __name__ == "__main__":
    demo.launch(server_name="0.0.0.0", server_port=7860)