app.py

import os
import cv2
import qrcode
import numpy as np
import gradio as gr
import mediapipe as mp
import tensorflow as tf
from PIL import Image, ImageOps

# --- 1. Load the Pre-trained TFLite Model ---
# We assume 'model.tflite' is uploaded alongside this script
interpreter = tf.lite.Interpreter(model_path="model.tflite")
interpreter.allocate_tensors()

# Get input and output details
input_details = interpreter.get_input_details()
output_details = interpreter.get_output_details()

print("Interpreter is ready for edge predictions!")

# --- 2. Setup MediaPipe for Background Removal ---
mp_selfie_segmentation = mp.solutions.selfie_segmentation
segmenter = mp_selfie_segmentation.SelfieSegmentation(model_selection=0)

def remove_background(image_pil):
    """
    Uses MediaPipe to segment the person from the background and replaces the background with white.
    """
    # Convert PIL image to numpy array (RGB)
    image_np = np.array(image_pil.convert("RGB"), copy=True)

    # Get the segmentation mask
    results = segmenter.process(image_np)
    mask = results.segmentation_mask

    # Create a condition: If probability > 0.5, it's a hand/person
    condition = np.stack((mask,) * 3, axis=-1) > 0.5

    # Create a solid white background
    bg_image = np.ones(image_np.shape, dtype=np.uint8) * 255

    # Combine: Keep original if condition is true, else use white background
    output_image = np.where(condition, image_np, bg_image)

    # Return as PIL Image
    return Image.fromarray(output_image).convert("L")

# --- 3. Prediction Logic ---
def predict_rps(image):
    if image is None:
        return None, None

    # STEP 1: Remove Background
    clean_image = remove_background(image)

    # STEP 2: Smart Crop & Resize to 300x300
    processed_image = ImageOps.fit(clean_image, (300, 300), Image.Resampling.LANCZOS)

    # STEP 3: Prepare for Model
    input_data = np.array(processed_image, dtype=np.float32)
    input_data = np.expand_dims(input_data, axis=-1) / 255.0
    input_data = np.expand_dims(input_data, axis=0)

    # STEP 4: Inference using Interpreter
    interpreter.set_tensor(input_details[0]['index'], input_data)
    interpreter.invoke()
    output_data = interpreter.get_tensor(output_details[0]['index'])

    # STEP 5: Format Results
    class_names = ['Rock ✊', 'Paper ✋', 'Scissors ✌️']
    prediction_dict = {class_names[i]: float(output_data[0][i]) for i in range(3)}

    return processed_image, prediction_dict

# --- 4. QR Code Helper ---
def generate_qr_code(url):
    if not url:
        return None
    qr = qrcode.QRCode(version=1, box_size=10, border=4)
    qr.add_data(url)
    qr.make(fit=True)
    return qr.make_image(fill_color="black", back_color="white").get_image()

# --- 5. Gradio UI ---
with gr.Blocks(theme="default", title="Edge AI: ✊Rock - ✋Paper - ✌️Scissors") as demo:

    gr.Markdown("# Edge AI: ✊Rock - ✋Paper - ✌️Scissors")
    gr.Markdown("This model is running on TFLite! Take a photo of your hand.")

    with gr.Row():
        with gr.Column():
            input_img = gr.Image(
                type="pil",
                label="Snap a Photo",
                sources=["webcam", "upload"]
            )
            predict_btn = gr.Button("Predict", variant="primary")

        with gr.Column():
            output_img = gr.Image(label="What the AI Sees")
            output_label = gr.Label(num_top_classes=3, label="Prediction")

    predict_btn.click(
        fn=predict_rps,
        inputs=input_img,
        outputs=[output_img, output_label]
    )

    # QR Code section
    gr.Markdown("---")
    with gr.Row():
        qr_display = gr.Image(label="Share this app", show_download_button=False, height=300, width=300)
    
    url_state = gr.Textbox(visible=False)
    
    # In Hugging Face Spaces, we just want to display the URL. 
    # The JS injection might behave differently inside an iframe, but we'll keep it simple.
    demo.load(
        fn=generate_qr_code,
        inputs=[url_state],
        outputs=[qr_display],
        js="() => window.location.href"
    )

if __name__ == "__main__":
    demo.launch()