app.py

import pandas as pd
import numpy as np
import tensorflow as tf
from keras.models import load_model
import gradio as gr

# Extended classes and labels
classes = [
    'car', 'house', 'wine bottle', 'chair', 'table',
    'tree', 'camera', 'fish', 'rain', 'clock', 'hat',
    'dog', 'cat', 'bicycle', 'plane', 'book', 'computer'
]

labels = {name: index for index, name in enumerate(classes)}
num_classes = len(classes)

# Load the model
model = load_model('sketch_recognition_model_cnn.h5')

# Predict function for interface
def predict_fn(image):
    """
    Predict the class of a drawn image.

    Args:
        image: The input image drawn by the user.

    Returns:
        The predicted class name.
    """
    try:
        # Preprocessing the image
        resized_image = tf.image.resize(image, (28, 28))  # Resize image to (28, 28)
        grayscale_image = tf.image.rgb_to_grayscale(resized_image)  # Convert image to grayscale
        image_array = np.array(grayscale_image) / 255.0  # Normalize the image

        # Prepare image for model input
        image_array = image_array.reshape(1, 28, 28, 1)  # Add batch dimension
        predictions = model.predict(image_array).reshape(num_classes)  # 2D output to 1D

        # Predict the class index
        predicted_index = tf.argmax(predictions).numpy()  # Get the index of the highest score
        class_name = classes[predicted_index]  # Retrieve the class name

        return class_name
    except Exception as e:
        return f"Error in prediction: {str(e)}"

# Gradio application interface
gr.Interface(
    fn=predict_fn,
    inputs="paint",
    outputs="label",
    title="DoodleDecoder",
    description="Draw something from: Car, House, Wine bottle, Chair, Table, Tree, Camera, Fish, Rain, Clock, Hat, Dog, Cat, Bicycle, Plane, Book, Computer",
    interpretation='default',
    article="Draw large with thick stroke."
).launch()