import gradio as gr import tensorflow as tf import numpy as np import cv2 model = tf.keras.models.load_model('my_model (1).keras') def center_and_resize(img): # 1. Find the bounding box of the digit coords = cv2.findNonZero(img) x, y, w, h = cv2.boundingRect(coords) # 2. Extract and create a square frame with padding digit = img[y:y+h, x:x+w] scale = 20.0 / max(w, h) new_w, new_h = int(w * scale), int(h * scale) digit = cv2.resize(digit, (new_w, new_h)) # 3. Paste into the center of a 28x28 black canvas canvas = np.zeros((28, 28), dtype=np.uint8) dx, dy = (28 - new_w) // 2, (28 - new_h) // 2 canvas[dy:dy+new_h, dx:dx+new_w] = digit return canvas def predict_digit(sketchpad_data): # Extract the drawing layer img = sketchpad_data['composite'] # Preprocess: Grayscale -> Invert -> Threshold -> Center img = cv2.cvtColor(img, cv2.COLOR_RGBA2GRAY) _, img = cv2.threshold(img, 50, 255, cv2.THRESH_BINARY) # If the canvas is empty, return zeros if cv2.countNonZero(img) == 0: return {str(i): 0.0 for i in range(10)} img = center_and_resize(img) # Normalize and prepare for model img = img.astype('float32') / 255.0 img = np.expand_dims(img, axis=(0, -1)) # Predict prediction = model.predict(img) return {str(i): float(prediction[0][i]) for i in range(10)} demo = gr.Interface( fn=predict_digit, inputs=gr.Sketchpad(label="Draw a digit", type="numpy"), outputs=gr.Label(num_top_classes=3), title="Robust MNIST Classifier" ) demo.launch()