app.py

import tensorflow as tf
import gradio as gr
import numpy as np
import cv2

# Load model
model = tf.keras.models.load_model('TP_MNIST_CNN_model.h5')

def preprocess_and_predict(input_data):
    if input_data is None:
        return None
        
    # Gradio 4.x Sketchpad returns a dict usually: {'composite': array, 'layers': [...]}
    # We take the composite image
    if isinstance(input_data, dict):
        image = input_data['composite']
    else:
        image = input_data

    # 1. Resize to MNIST standard (28x28)
    # Interpolation AREA is better for shrinking images without losing thin lines
    image = cv2.resize(image, (28, 28), interpolation=cv2.INTER_AREA)

    # 2. Convert to Grayscale (if RGB/RGBA)
    if len(image.shape) == 3:
        image = cv2.cvtColor(image, cv2.COLOR_RGB2GRAY)

    # 3. Normalize & Invert Logic
    # MNIST is White digits on Black background.
    # If user draws Black on White, we must invert.
    # Check mean pixel intensity: if high (>127), background is likely white.
    if np.mean(image) > 127:
        image = 255 - image
        
    # 4. Normalize 0-1
    image = image / 255.0

    # 5. Reshape for Model (Batch, Height, Width, Channels)
    image = image.reshape(1, 28, 28, 1)

    # Prediction
    predictions = model.predict(image, verbose=0)
    return int(np.argmax(predictions))

# Modern Gradio 4 Interface
iface = gr.Interface(
    fn=preprocess_and_predict,
    inputs=gr.Sketchpad(label="Dessinez un chiffre", type="numpy"),
    outputs="label",
    title="MNIST Digit Recognition",
    description="Reconnaissance de chiffres manuscrits via CNN. Dessinez au centre.",
    allow_flagging="never"
)

iface.launch()