Lab on Python

digitRecognizer.py

# Handwritten digit recognition for MNIST dataset using Convolutional Neural Networks # Step 1: Import all required keras libraries from keras.datasets import mnist # This is used to load mnist dataset later from keras.utils import np_utils # This will be used to convert your test image to a categorical class (digit from 0 to 9) # Step 2: Load and return training and test datasets def load_dataset(): # 2a. Load dataset X_train, X_test, y_train, y_test via imported keras library # 2b. reshape for X train and test vars - Hint: X_train = X_train.reshape((X_train.shape[0], 28, 28, 1)).astype('float32') # 2c. normalize inputs from 0-255 to 0-1 - Hint: X_train = X_train / 255 # 2d. Convert y_train and y_test to categorical classes - Hint: y_train = np_utils.to_categorical(y_train) # 2e. return your X_train, X_test, y_train, y_test # Step 3: define your CNN model here in this function and then later use this function to create your model def digit_recognition_cnn(): # 3a. create your CNN model here with Conv + ReLU + Flatten + Dense layers # 3b. Compile your model with categorical_crossentropy (loss), adam optimizer and accuracy as a metric # 3c. return your model # Step 4: Call digit_recognition_cnn() to build your model # Step 5: Train your model and see the result in Command window. Set epochs to a number between 10 - 20 and batch_size between 150 - 200 # Step 6: Evaluate your model via your_model_name.evaluate() function and copy the result in your report # Step 7: Save your model via your_model_name.save('digitRecognizer.h5') # Code below to make a prediction for a new image. # Step 8: load required keras libraries from keras.preprocessing.image import load_img from keras.preprocessing.image import img_to_array from keras.models import load_model # Step 9: load and normalize new image def load_new_image(path): # 9a. load new image newImage = load_img(path, grayscale=True, target_size=(28, 28)) # 9b. Convert image to array newImage = img_to_array(newImage) # 9c. reshape into a single sample with 1 channel (similar to how you reshaped in load_dataset function) # 9d. normalize image data - Hint: newImage = newImage / 255 # 9e. return newImage # Step 10: load a new image and predict its class def test_model_performance(): # 10a. Call the above load image function img = load_new_image('your_new_image_file_path') # 10b. load your CNN model (digitRecognizer.h5 file) # 10c. predict the class - Hint: imageClass = your_model_name.predict_classes(img) # 10d. Print prediction result print(imageClass[0]) # Step 11: Test model performance here by calling the above test_model_performance function