This is taken from the book “Deep Learning with Python” by François Chollet. This book is a good starter point for learning Machine Learning: https://www.manning.com/books/deep-learning-with-python.
Download MNIST dataset
from keras.datasets import mnist
(train_images, train_labels), (test_images, test_labels) = mnist.load_data()
Prepare a network using Sequential model with 2 fully connected or Dense layers of neurons. The first layer with 512 nodes with relu activation function. The second layer with 10 nodes, corresponding to 10 labels and using softmax activation. The output of this layer will be an array of ten floats, adding up to 1, each represents the probability of the sample to be classified as the corresponding digit.
from keras import models
from keras import layers
network = models.Sequential()
network.add(layers.Dense(512, activation='relu', input_shape=(28 * 28,)))
network.add(layers.Dense(10, activation='softmax'))
Here we compile the neural net by specifying the loss function, the optimizer and the evaluation metric.
network.compile(optimizer='rmsprop',
loss='categorical_crossentropy',
metrics=['accuracy'])
Normalize the train and test data. The shade of each pixel is represented by a float32 from 0 to 1, instead of an uint_8 from 0 to 255.
train_images = train_images.reshape((60000, 28 * 28))
train_images = train_images.astype('float32') / 255
test_images = test_images.reshape((10000, 28 * 28))
test_images = test_images.astype('float32') / 255
Here, we encode the labels categorically.
from keras.utils import to_categorical
train_labels = to_categorical(train_labels)
test_labels = to_categorical(test_labels)
Now we train the neural net. We will do this training in 5 epochs and a batch_size of 128. Meaning we update the weights after every 128 samples, and we will iterate through the whole training set 5 times. Two quantities are displayed during training: the loss of the network over the training data, and the accuracy of the network over the training data.
network.fit(train_images, train_labels, epochs=5, batch_size=128)
Epoch 1/5
469/469 [==============================] - 1s 1ms/step - loss: 0.4275 - accuracy: 0.8778
Epoch 2/5
469/469 [==============================] - 0s 982us/step - loss: 0.1145 - accuracy: 0.9665
Epoch 3/5
469/469 [==============================] - 0s 960us/step - loss: 0.0692 - accuracy: 0.9795
Epoch 4/5
469/469 [==============================] - 0s 965us/step - loss: 0.0485 - accuracy: 0.9852
Epoch 5/5
469/469 [==============================] - 0s 991us/step - loss: 0.0389 - accuracy: 0.9892
<tensorflow.python.keras.callbacks.History at 0x7f6543f8f4a8>
After the training finishes, we can evaluate the performance of the network with a test dataset.
test_loss, test_acc = network.evaluate(test_images, test_labels)
print(f"test_acc: {test_acc}")
313/313 [==============================] - 0s 676us/step - loss: 0.0667 - accuracy: 0.9790
test_acc: 0.9789999723434448