pdfcoffee

Generative Adversarial Networks
DCGAN for MNIST digits
Let us now build a DCGAN for generating handwritten digits. We first see the
code for the generator. The generator is built by adding the layers sequentially. The
first layer is a dense layer that takes the noise of 100 dimensions as an input. The
100-dimensional input is expanded to a flat vector of size 128 × 7 × 7. This is done so
that finally we get an output of size 28 × 28, the standard size of MNIST handwritten
digits. The vector is reshaped to a tensor of size 7 × 7 × 128. This vector is then
upsampled using TensorFlow Keras UpSampling2D layer. Please note that this layer
simply scales up the image by doubling rows and columns. The layer has no weights,
so it is computationally cheap.
The Upsampling2D layer will now double the rows and columns of the 7 × 7 × 128
(rows × columns × channels) image, yielding an output of size 14 × 14 × 128. The
upsampled image is passed to a convolutional layer. This convolutional layer learns
to fill in the details in the upsampled image. The output of convolution is passed
to batch normalization for better gradient flow. The batch normalized output then
undergoes ReLU activation in all the intermediate layers. We repeat the structure,
that is, upsampling | convolution | batch normalization | ReLU. In the following
generator we have two such structures, the first with 128 filters, and the second
with 64 filters in the convolution operation. The final output is obtained from
a pure convolutional layer with 3 filters and tan hyperbolic activation, yielding
an image of size 28 × 28 × 1:
def build_generator(self):
model = Sequential()
model.add(Dense(128 * 7 * 7, activation="relu", input_dim=self.
latent_dim))
model.add(Reshape((7, 7, 128)))
model.add(UpSampling2D())
model.add(Conv2D(128, kernel_size=3, padding="same"))
model.add(BatchNormalization(momentum=0.8))
model.add(Activation("relu"))
model.add(UpSampling2D())
model.add(Conv2D(64, kernel_size=3, padding="same"))
model.add(BatchNormalization(momentum=0.8))
model.add(Activation("relu"))
model.add(Conv2D(self.channels, kernel_size=3, padding="same"))
model.add(Activation("tanh"))
model.summary()
noise = Input(shape=(self.latent_dim,))
img = model(noise)
return Model(noise, img)
[ 200 ]