import gradio as gr
import tensorflow as tf
import numpy as np
from PIL import Image
model_path = "BMW_transfer_learning.keras"
model = tf.keras.models.load_model(model_path)
# Define the core prediction function
def predict_bmw(image):
    # Preprocess image
    print(type(image))
    image = Image.fromarray(image.astype('uint8'))  # Convert numpy array to PIL image
    image = image.resize((150, 150)) #resize the image to 28x28 and converts it to gray scale
    image = np.array(image)
    image = np.expand_dims(image, axis=0) # same as image[None, ...]
    
    # Predict
    prediction = model.predict(image)

    # Because the output layer was dense(0) without an activation function, we need to apply sigmoid to get the probability
    # we could also change the output layer to dense(4, activation='sigmoid')
    prediction = np.round(prediction, 2)
    
    # Separate the probabilities for each class
    p_m2 = prediction[0][0]  # Probability for class 'm2'
    p_m4 = prediction[0][1]   # Probability for class 'm4'
    p_x6 = prediction[0][2]    # Probability for class 'x6'
    p_i8 = prediction[0][3]    # Probability for class 'i8'

    

    return {'M2': p_m2, 'M4': p_m4, 'X6': p_x6, 'i8': p_i8}

    
# Create the Gradio interface
input_image = gr.Image()
iface = gr.Interface(
    fn=predict_bmw,
    inputs=input_image, 
    outputs=gr.Label(),
    examples=["images/m2.jpg", "images/m2_1.jpg", "images/m4.jpg", "images/m4_1.jpg", "images/x6.jpg", "images/x6_1.jpg", "images/i8.jpg", "images/i8_1.jpg"],   
    description="A simple mlp classification model for image classification using the mnist dataset.")
iface.launch(share=True)