def generator(z, bs, is_training=True, reuse=None):
with tf.variable_scope("generator", reuse=reuse) as scope_g:
# Variables for classifier
G_B1 = utils.bias_variable([3], name="G_B1")
G_B2 = utils.bias_variable([K], name="G_B2")
G_B3 = utils.bias_variable([8 * 8 * L], name="G_B3")
G_B4 = utils.bias_variable([M], name="G_B4")
GH3 = tf.nn.relu(
utils.bn((tf.matmul(z, tf.transpose(C_W4)) + G_B4),
is_training=is_training,
scope="G_bn_gh3"))
GH2 = tf.nn.relu(
utils.bn((tf.matmul(GH3, tf.transpose(C_W3)) + G_B3),
is_training=is_training,
scope="G_bn_gh2"))
GHH2 = tf.reshape(GH2, shape=[-1, 8, 8, L])
stride = 2 # output is 14x14
GH1 = tf.nn.relu(
tf.nn.conv2d_transpose(GHH2,
C_W2,
output_shape=[bs, 16, 16, K],
strides=[1, stride, stride, 1]) +
G_B2) #deconv2 W2
stride = 2 # output is 28x28
GXlogits = tf.nn.conv2d_transpose(GH1,
C_W1,
output_shape=[bs, 32, 32, 3],
strides=[1, stride, stride, 1
]) + G_B1 #deconv2 W1
GXsigmoid = tf.nn.sigmoid(GXlogits)
return GXsigmoid, GXlogits
def classifier(x, is_training=True, reuse=None):
with tf.variable_scope("classifier", reuse=reuse) as scope_c:
# Variables for classifier
C_B1 = utils.bias_variable([K], name="C_B1")
C_B2 = utils.bias_variable([L], name="C_B2")
C_B3 = utils.bias_variable([M], name="C_B3")
C_B4 = utils.bias_variable([num_classes], name="C_B4")
stride = 2 # output is 16x16
H1 = lrelu(
tf.nn.conv2d(
x, C_W1, strides=[1, stride, stride, 1], padding='SAME') +
C_B1)
stride = 2 # output is 8x8
H2 = lrelu(
utils.bn((tf.nn.conv2d(
H1, C_W2, strides=[1, stride, stride, 1], padding='SAME') +
C_B2),
is_training=is_training,
scope="C_bn_h2"))
# reshape the output from the third convolution for the fully connected layer
HH2 = tf.reshape(H2, shape=[-1, 8 * 8 * L])
H3 = tf.nn.relu(tf.matmul(HH2, C_W3) + C_B3)
Ylogits = tf.matmul(H3, C_W4) + C_B4
Ysigmoid = tf.nn.sigmoid(Ylogits)
Ysoftmax = tf.nn.softmax(Ylogits)
return Ysoftmax, Ysigmoid, Ylogits
def discriminator(x, is_training=True, reuse=None):
with tf.variable_scope("discriminator", reuse=reuse) as scope:
# Variables for classifier
D_W1 = utils.weight_variable([4, 4, 3, K], name="D_W1")
D_B1 = utils.bias_variable([K], name="D_B1")
D_W2 = utils.weight_variable([4, 4, K, L], name="D_W2")
D_B2 = utils.bias_variable([L], name="D_B2")
D_W3 = utils.weight_variable([8 * 8 * L, M], name="D_W3")
D_B3 = utils.bias_variable([M], name="D_B3")
D_W4 = utils.weight_variable([M, 1], name="D_W4")
D_B4 = utils.bias_variable([1], name="D_B4")
stride = 2 # output is 16x16
H1 = lrelu(
tf.nn.conv2d(
x, D_W1, strides=[1, stride, stride, 1], padding='SAME') +
D_B1)
print(H1.shape)
stride = 2 # output is 8x8
H2 = lrelu(
utils.bn((tf.nn.conv2d(
H1, D_W2, strides=[1, stride, stride, 1], padding='SAME') +
D_B2),
is_training=is_training,
scope="D_bn_h2"))
print(H2.shape)
# reshape the output from the third convolution for the fully connected layer
HH2 = tf.reshape(H2, shape=[-1, 8 * 8 * L])
H3 = lrelu(tf.matmul(HH2, D_W3) + D_B3)
Ylogits = tf.matmul(H3, D_W4) + D_B4
Ysigmoid = tf.nn.sigmoid(Ylogits)
Ysoftmax = tf.nn.softmax(Ylogits)
return Ysoftmax, Ysigmoid, Ylogits