import gradio as gr
import tensorflow as tf
import numpy as np
from PIL import ImageOps

# Load your trained model
model = tf.keras.models.load_model("hand_sign_model.h5")

# Sign Language MNIST class names (A-Z except J and Z)
class_names = [chr(i) for i in range(65, 91) if i not in [74, 90]]

def preprocess_image(image, target_size=(28, 28)):
    """
    Preprocess input PIL image:
    - Convert to grayscale
    - Resize to target_size
    - Normalize pixel values to [0,1]
    - Add batch and channel dimensions for model input
    """
    # Convert to grayscale
    image = ImageOps.grayscale(image)
    
    # Resize to target size expected by the model
    image = image.resize(target_size)
    
    # Convert to numpy array and normalize
    img_array = np.array(image) / 255.0
    
    # Add batch and channel dimensions
    img_array = np.expand_dims(img_array, axis=(0, -1))
    return img_array

def predict(image):
    if image is None:
        return "Please upload an image."
    
    # Preprocess the image to model input shape
    img_array = preprocess_image(image, target_size=(28, 28))
    
    # Predict probabilities for each class
    prediction = model.predict(img_array)[0]
    
    # Get predicted class label and confidence
    predicted_label = class_names[np.argmax(prediction)]
    confidence = np.max(prediction) * 100
    
    return f"Predicted Sign: {predicted_label} ({confidence:.2f}%)"

# Build Gradio interface
demo = gr.Interface(
    fn=predict,
    inputs=gr.Image(type="pil"),
    outputs="text",
    title="Hand Sign Recognition",
    description="Upload any hand sign image of any size, and the model will predict the sign."
)

# Launch the app
demo.launch()