pdfcoffee

Chapter 4
Where pool_size=(2, 2) is a tuple of two integers representing the factors by
which the image is vertically and horizontally downscaled. So (2, 2) will halve the
image in each dimension, and strides=(2, 2) is the stride used for processing.
Now, let us review the code. First, we import a number of modules:
import tensorflow as tf
from tensorflow.keras import datasets, layers, models, optimizers
# network and training
EPOCHS = 5
BATCH_SIZE = 128
VERBOSE = 1
OPTIMIZER = tf.keras.optimizers.Adam()
VALIDATION_SPLIT=0.95
IMG_ROWS, IMG_COLS = 28, 28 # input image dimensions
INPUT_SHAPE = (IMG_ROWS, IM G_COLS, 1)
NB_CLASSES = 10 # number of outputs = number of digits
Then we define the LeNet network:
#define the convnet
def build(input_shape, classes):
model = models.Sequential()
We have a first convolutional stage with rectified linear unit (ReLU) activations
followed by a max pooling. Our net will learn 20 convolutional filters, each one of
which with a size of 5×5. The output dimension is the same as the input shape, so it
will be 28×28. Note that since Convolution2D is the first stage of our pipeline, we
are also required to define its input_shape. The max pooling operation implements
a sliding window that slides over the layer and takes the maximum of each region
with a step of 2 pixels both vertically and horizontally:
# CONV => RELU => POOL
model.add(layers.Convolution2D(20, (5, 5), activation='relu', input_
shape=input_shape))
model.add(layers.MaxPooling2D(pool_size=(2, 2), strides=(2, 2)))
Then there is a second convolutional stage with ReLU activations, followed again
by a max pooling layer. In this case we increase the number of convolutional filters
learned to 50 from the previous 20. Increasing the number of filters in deeper layers
is a common technique used in deep learning:
# CONV => RELU => POOL
model.add(layers.Convolution2D(50, (5, 5), activation='relu'))
model.add(layers.MaxPooling2D(pool_size=(2, 2), strides=(2, 2)))
[ 115 ]