import gradio as gr
import tensorflow as tf
from tensorflow.keras.models import load_model
import numpy as np
from PIL import Image

# Load the pre-trained model (assumed trained on Sign Language MNIST)
model = load_model('sign_language_mnist_cnn.h5')

# Define class names (A-Z, skipping J=9 and Z=25 due to dataset constraints)
CLASS_NAMES = list('ABCDEFGHIKLMNOPQRSTUVWXY')  # 24 classes (0-8, 10-24)

# Preprocessing function for Sign Language MNIST
def preprocess_image(image: Image.Image):
    # Convert to grayscale
    image = image.convert('L')  # 'L' mode for grayscale
    
    # Resize to 28x28 (matching dataset)
    image = image.resize((28, 28))
    
    # Convert to numpy array and normalize to 0-255 range (as in dataset)
    image_array = np.array(image)
    
    # Normalize to 0-1 range (common for model input)
    image_array = image_array / 255.0
    
    # Add batch and channel dimensions (1, 28, 28, 1)
    image_array = np.expand_dims(image_array, axis=(0, -1))
    
    return image_array

# Prediction function
def predict_sign(image):
    processed_image = preprocess_image(image)
    
    # Get model predictions (logits)
    predictions = model.predict(processed_image)
    probability=np.max(predictions)
    target=['A','B','C','D','E','F','G','H','I','K','L','M','N','O','P','Q','R','S','T','U','V','W','X','Y']
    arg=np.argmax(predictions)
    result={'prediction':target[arg],'probability':probability}

    return result

# Gradio interface
interface = gr.Interface(
    fn=predict_sign,
    inputs=gr.Image(type="pil", label="Upload a Hand Gesture Image"),
    outputs=gr.Textbox(label="Prediction"),
    title="Sign Language MNIST Classifier",
    description="Upload an image of a hand gesture to classify it as a letter (A-Z, excluding J and Z)."
)

# Launch the app
interface.launch()