Convolutional neural networks (CNNs) in `keras`

Lots of keras examples, some including CNNs available here: https://github.com/fchollet/keras/tree/master/examples

Specifically, this notebook is based on the following example training a CNN on the MNIST dataset of hand-written digits: https://github.com/fchollet/keras/blob/master/examples/mnist_cnn.py



In [1]:

    
import matplotlib.cm as cm
import matplotlib.pyplot as plt
import tensorflow.contrib.keras as keras



In [2]:

    
%matplotlib inline

Dataset pre-processing



In [3]:

    
# Dataset of 60,000 28x28 grayscale images of the 10 digits, along with a test set of 10,000 images.
(x_train, y_train), (x_test, y_test) = keras.datasets.mnist.load_data()
# input image dimensions and class counts
img_rows, img_cols = 28, 28
num_classes = 10



In [4]:

    
x_train[0].shape









    Out[4]:





(28, 28)



In [5]:

    
y_train[0]









    Out[5]:





5



In [6]:

    
plt.imshow(x_train[0], cmap=cm.binary)









    Out[6]:





<matplotlib.image.AxesImage at 0x1229dc9e8>



In [7]:

    
# images are expected as 3D tensors with the third dimension containing different image channels; reshape x to a
# 3D tensore with single color channel, the grayscale channel
x_train = x_train.reshape(x_train.shape[0], img_rows, img_cols, 1)
x_test = x_test.reshape(x_test.shape[0], img_rows, img_cols, 1)
input_shape = (img_rows, img_cols, 1)
x_train[0].shape









    Out[7]:





(28, 28, 1)



In [8]:

    
# convert X to [0,1]
x_train = x_train.astype('float32')
x_test = x_test.astype('float32')
x_train /= 255
x_test /= 255
print('x_train shape:', x_train.shape)
print(x_train.shape[0], 'train samples')
print(x_test.shape[0], 'test samples')









    



x_train shape: (60000, 28, 28, 1)
60000 train samples
10000 test samples



In [9]:

    
y_train[:5]









    Out[9]:





array([5, 0, 4, 1, 9], dtype=uint8)



In [10]:

    
# convert to a one hot encoding of the class labels
y_train = keras.utils.to_categorical(y_train, num_classes)
y_test = keras.utils.to_categorical(y_test, num_classes)
y_train[:5]









    Out[10]:





array([[ 0.,  0.,  0.,  0.,  0.,  1.,  0.,  0.,  0.,  0.],
       [ 1.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.],
       [ 0.,  0.,  0.,  0.,  1.,  0.,  0.,  0.,  0.,  0.],
       [ 0.,  1.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.],
       [ 0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  1.]])

Specifying the CNN model



In [11]:

    
batch_size = 128
epochs = 7  # increasing this would probably make sense but takes longer to compute



In [12]:

    
# houses a linear stack of layers
model = keras.models.Sequential()



In [13]:

    
# add layers to the sequential model
model.add(keras.layers.Conv2D(32, # 32 filters/kernels
                 kernel_size=(3, 3), # filter size of 3x3 pixels
                 activation='relu', 
                 input_shape=input_shape))
model.add(keras.layers.Conv2D(64, (3, 3), activation='relu'))
model.add(keras.layers.MaxPooling2D(pool_size=(2, 2)))
model.add(keras.layers.Dropout(0.25))
model.add(keras.layers.Flatten())
model.add(keras.layers.Dense(128, activation='relu'))
model.add(keras.layers.Dropout(0.5))
model.add(keras.layers.Dense(num_classes, activation='softmax'))

model.compile(loss=keras.losses.categorical_crossentropy,
              optimizer=keras.optimizers.Adadelta(),
              metrics=['accuracy'])

The model can be visualized as follows:



In [14]:

    
keras.utils.plot_model(model, to_file='chapter_9_cnn.png', show_shapes=True)

A convolutional layer 'Conv2D' lookse like this:

A max pooling layer 'MaxPooling2D' lookse like this:

Training the model



In [15]:

    
model.fit(x_train, y_train,
          batch_size=batch_size,
          epochs=epochs,
          verbose=1,
          validation_data=(x_test, y_test))









    



Train on 60000 samples, validate on 10000 samples
Epoch 1/7
60000/60000 [==============================] - 131s - loss: 0.3785 - acc: 0.8838 - val_loss: 0.0904 - val_acc: 0.9708
Epoch 2/7
60000/60000 [==============================] - 130s - loss: 0.1409 - acc: 0.9581 - val_loss: 0.0611 - val_acc: 0.9804
Epoch 3/7
60000/60000 [==============================] - 124s - loss: 0.1079 - acc: 0.9680 - val_loss: 0.0516 - val_acc: 0.9832
Epoch 4/7
60000/60000 [==============================] - 128s - loss: 0.0929 - acc: 0.9731 - val_loss: 0.0457 - val_acc: 0.9858
Epoch 5/7
60000/60000 [==============================] - 128s - loss: 0.0820 - acc: 0.9756 - val_loss: 0.0396 - val_acc: 0.9873
Epoch 6/7
60000/60000 [==============================] - 129s - loss: 0.0725 - acc: 0.9785 - val_loss: 0.0359 - val_acc: 0.9883
Epoch 7/7
60000/60000 [==============================] - 127s - loss: 0.0680 - acc: 0.9803 - val_loss: 0.0347 - val_acc: 0.9887






    Out[15]:





<tensorflow.contrib.keras.python.keras.callbacks.History at 0x131c02978>

Testing the model



In [16]:

    
score = model.evaluate(x_test, y_test, verbose=0)
print('Test loss:', score[0])
print('Test accuracy:', score[1])









    



Test loss: 0.0346899191732
Test accuracy: 0.9887

Convolutional neural networks (CNNs) in keras

Dataset pre-processing

Specifying the CNN model

Training the model

Testing the model

Convolutional neural networks (CNNs) in `keras`