Chapter 2.1 - A first look at a neural network

In [1]:

    

from keras.datasets import mnist


    

    




Using TensorFlow backend.

In [2]:

    

(train_images, train_labels),(test_images, test_labels) = mnist.load_data()


    

In [3]:

    

train_images.shape


    

    
Out[3]:





(60000, 28, 28)

In [4]:

    

train_labels.shape


    

    
Out[4]:





(60000,)

In [5]:

    

train_labels


    

    
Out[5]:





array([5, 0, 4, ..., 5, 6, 8], dtype=uint8)

In [6]:

    

test_images.shape


    

    
Out[6]:





(10000, 28, 28)

In [7]:

    

test_labels.shape


    

    
Out[7]:





(10000,)

In [8]:

    

test_labels


    

    
Out[8]:





array([7, 2, 1, ..., 4, 5, 6], dtype=uint8)

In [9]:

    

# Importing models like Sequential or Functional Model
from keras import models


    

In [10]:

    

# Imporing layers like Dense
from keras import layers


    

In [11]:

    

# Initializing a sequential network
network = models.Sequential()


    

In [12]:

    

# Adding the input layer
network.add(layers.Dense(units = 512, 
                         activation = 'relu', 
                         input_shape = (28 * 28,)))


    

In [13]:

    

# Adding the outpul layer
network.add(layers.Dense(units = 10, 
                         activation = 'softmax'))


    

In [14]:

    

# Compiling the network
network.compile(optimizer='rmsprop',
                loss='categorical_crossentropy',
                metrics = ['accuracy'])


    

In [15]:

    

# Reshaping the data from a 3-D array (number of 2-d examples) to a 2-D array (number of 1-D examples)
train_images = train_images.reshape((60000, 28 * 28))


    

In [16]:

    

# Changing type of the data to float32
train_images = train_images.astype('float32')


    

In [17]:

    

# Rescaling the data from 0-255 to 0-1
train_images = train_images / 255.0


    

In [18]:

    

test_images = test_images.reshape((10000, 28 * 28))
test_images = test_images.astype('float32')
test_images = test_images / 255.0


    

In [19]:

    

# Importing one-hot encoding tool to_categorical from keras.utils
from keras.utils import to_categorical


    

In [20]:

    

# One-hot encoding the labels
train_labels = to_categorical(train_labels)
test_labels = to_categorical(test_labels)


    

In [21]:

    

network.fit(train_images, 
            train_labels, 
            epochs = 5,
            batch_size = 128)


    

    




Epoch 1/5
60000/60000 [==============================] - 3s - loss: 0.2591 - acc: 0.9242     
Epoch 2/5
60000/60000 [==============================] - 2s - loss: 0.1031 - acc: 0.9693     
Epoch 3/5
60000/60000 [==============================] - 2s - loss: 0.0683 - acc: 0.9799     
Epoch 4/5
60000/60000 [==============================] - 2s - loss: 0.0502 - acc: 0.9851     
Epoch 5/5
60000/60000 [==============================] - 2s - loss: 0.0381 - acc: 0.9883     






    
Out[21]:





<keras.callbacks.History at 0x27a3bbd5c18>

In [22]:

    

test_loss, test_acc = network.evaluate(test_images, test_labels)


    

    




 9472/10000 [===========================>..] - ETA: 0s

In [23]:

    

print('Test acc:', test_acc)


    

    




Test acc: 0.9789