def generator(z, y, image_size, batch_size, y_dim, gfc_dim, gf_dim, c_dim):
with tf.variable_scope("generator"):
g_bn0 = BatchNorm(name='g_bn0')
g_bn1 = BatchNorm(name='g_bn1')
g_bn2 = BatchNorm(name='g_bn2')
# taken from https://github.com/carpedm20/DCGAN-tensorflow/blob/master/model.py
s_h, s_w = image_size[1], image_size[0]
s_h2, s_h4 = int(s_h / 2), int(s_h / 4)
s_w2, s_w4 = int(s_w / 2), int(s_w / 4)
yb = tf.reshape(y, [batch_size, 1, 1, y_dim])
z = tf.concat([z, y], 1)
h0 = tf.nn.relu(g_bn0(linear(z, gfc_dim, 'g_h0_lin')))
h0 = tf.concat([h0, y], 1)
h1 = tf.nn.relu(
g_bn1(linear(h0, gf_dim * 2 * s_h4 * s_w4, 'g_h1_lin')))
h1 = tf.reshape(h1, [batch_size, s_h4, s_w4, gf_dim * 2])
h1 = conv_cond_concat(h1, yb)
h2 = tf.nn.relu(
g_bn2(
deconv2d(h1, [batch_size, s_h2, s_w2, gf_dim * 2],
name='g_h2')))
h2 = conv_cond_concat(h2, yb)
h3 = tf.nn.sigmoid(
deconv2d(h2, [batch_size, s_h, s_w, c_dim], name='g_h3'))
return tf.identity(h3, 'generator')
def discriminator(x,
y,
batch_size,
y_dim,
c_dim,
df_dim,
dfc_dim,
reuse=False):
with tf.variable_scope("discriminator") as scope:
if reuse:
scope.reuse_variables()
d_bn1 = BatchNorm(name='d_bn1')
d_bn2 = BatchNorm(name='d_bn2')
yb = tf.reshape(y, [batch_size, 1, 1, y_dim])
x = conv_cond_concat(x, yb)
h0 = lrelu(conv2d(x, c_dim + y_dim, name='d_h0_conv'))
h0 = conv_cond_concat(h0, yb)
h1 = lrelu(d_bn1(conv2d(h0, df_dim + y_dim, name='d_h1_conv')))
h1 = tf.reshape(h1, [batch_size, -1])
h1 = tf.concat([h1, y], 1)
h2 = lrelu(d_bn2(linear(h1, dfc_dim, 'd_h2_lin')))
h2 = tf.concat([h2, y], 1)
h3 = linear(h2, 1, 'd_h3_lin')
return tf.nn.sigmoid(h3), h3
def create_generator(self, z, y, scope_name, is_training=True, reuse=False):
with tf.variable_scope(scope_name) as scope:
if reuse:
scope.reuse_variables()
batch_norm_params = {
# 'decay': 0.999,
'decay': 0.9, # also known as momentum, they are the same
'updates_collections': None,
# 'epsilon': 0.001,
'epsilon': 1e-5,
'scale': True,
'is_training': is_training,
'scope': 'batch_norm',
}
# first argument is where to apply these
with arg_scope([layers.conv2d, layers.conv2d_transpose, layers.fully_connected],
normalizer_fn=layers.batch_norm,
normalizer_params=batch_norm_params,
weights_initializer=layers.xavier_initializer(uniform=False),
biases_initializer=tf.constant_initializer(0.0)
):
# taken from https://github.com/carpedm20/DCGAN-tensorflow/blob/master/model.py
s_h, s_w = self.image_size[1], self.image_size[0]
s_h2, s_h4 = int(s_h / 2), int(s_h / 4)
s_w2, s_w4 = int(s_w / 2), int(s_w / 4)
yb = tf.reshape(y, [self.batch_size, 1, 1, self.y_dim])
z = tf.concat([z, y], 1)
h0 = slim.fully_connected(z,
num_outputs=self.gfc_dim,
scope='g_h0_lin',
activation_fn=slim.nn.relu,
)
h0 = tf.concat([h0, y], 1)
h1 = slim.fully_connected(h0,
num_outputs=self.gf_dim * 2 * s_h4 * s_w4,
scope='g_h1_lin',
activation_fn=slim.nn.relu,
)
h1 = tf.reshape(h1, [self.batch_size, s_h4, s_w4, self.gf_dim * 2])
h1 = conv_cond_concat(h1, yb)
h2 = slim.conv2d_transpose(h1,
num_outputs=self.gf_dim * 2,
scope='g_h2',
kernel_size=[5, 5],
stride=2,
activation_fn=slim.nn.relu,
)
h2 = conv_cond_concat(h2, yb)
h3 = slim.conv2d_transpose(h2,
num_outputs=self.c_dim,
scope='g_h3',
kernel_size=[5, 5],
stride=2,
normalizer_fn=None,
activation_fn=slim.nn.sigmoid
)
return h3
def create_discriminator(self, x, y, scope_name, is_training=True, reuse=False):
with tf.variable_scope(scope_name) as scope:
if reuse:
scope.reuse_variables()
batch_norm_params = {
# 'decay': 0.999,
'decay': 0.9, # also known as momentum, they are the same
'updates_collections': None,
# 'epsilon': 0.001,
'epsilon': 1e-5,
'scale': True,
'is_training': is_training,
'scope': 'batch_norm',
}
with arg_scope([layers.conv2d, layers.conv2d_transpose, layers.fully_connected],
normalizer_fn=layers.batch_norm,
normalizer_params=batch_norm_params,
weights_initializer=layers.xavier_initializer(uniform=False),
biases_initializer=tf.constant_initializer(0.0)
):
yb = tf.reshape(y, [self.batch_size, 1, 1, self.y_dim])
x = conv_cond_concat(x, yb)
h0 = slim.conv2d(x,
num_outputs=self.c_dim + self.y_dim,
scope='d_h0_conv',
kernel_size=[5, 5],
stride=[2, 2],
normalizer_fn=None,
activation_fn=lrelu,
)
h0 = conv_cond_concat(h0, yb)
# not having bias variables here because of bias adding in batch normalization, see: https://stackoverflow.com/questions/46256747/can-not-use-both-bias-and-batch-normalization-in-convolution-layers
h1 = slim.conv2d(h0,
num_outputs=self.df_dim + self.y_dim,
scope='d_h1_conv',
kernel_size=[5, 5],
stride=[2, 2],
activation_fn=lrelu
)
h1 = tf.reshape(h1, [self.batch_size, -1])
h1 = tf.concat([h1, y], 1)
h2 = slim.fully_connected(h1,
num_outputs=self.dfc_dim,
scope='d_h2_lin',
activation_fn=lrelu,
)
h2 = tf.concat([h2, y], 1)
h3 = slim.fully_connected(h2,
num_outputs=1,
scope='d_h3_lin',
normalizer_fn=None,
activation_fn=None
)
return tf.nn.sigmoid(h3), h3