def _build_discriminator(self, tensor=None, training=False):
with tf.variable_scope('discriminator') as scope:
# set reuse if necessary
if tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES,
scope=scope.name):
scope.reuse_variables()
# discriminator base
tensor = self._build_discriminator_base(tensor, training)
# final layer
d_out = 2
with tf.variable_scope('final.{0}-{1}'.format(
tensor.shape[-1], d_out)):
out_logits = tf.reshape(
tf.layers.conv2d(tensor,
d_out,
4,
2,
'valid',
kernel_initializer=init_normal(),
name='conv'), [-1, d_out])
return tf.nn.softmax(out_logits), out_logits
def _build_generator(self, tensor=None, training=False, batch_norm=None):
assert self.n_pixel % 16 == 0, "isize has to be a multiple of 16"
nfilt = 2000
csize = 4
if tensor is None:
tensor = self.input_n
if batch_norm is None:
batch_norm = self.batch_norm_G
if batch_norm:
def bn(x, name=None):
return tf.contrib.layers.batch_norm(x,
is_training=training,
renorm=BATCH_RENORM,
decay=BATCH_NORM_DECAY)
else:
bn = tf.identity
with tf.variable_scope('generator') as scope:
# set reuse if necessary
if tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES,
scope=scope.name):
scope.reuse_variables()
# initial layer
with tf.variable_scope('initial.{0}-{1}'.format(
self.n_noise, nfilt)):
tensor = tf.nn.relu(
bn(tf.layers.conv2d_transpose(
tf.reshape(tensor, [-1, 1, 1, self.n_noise]),
nfilt,
4,
2,
'valid',
use_bias=not batch_norm,
kernel_initializer=init_normal(),
name='conv'),
name='bn'))
# upscaling layers
while csize < self.n_pixel // 2:
with tf.variable_scope('pyramid.{0}-{1}'.format(
nfilt, nfilt / 2)):
tensor = tf.nn.relu(
bn(tf.layers.conv2d_transpose(
tensor,
nfilt // 2,
4,
2,
'same',
use_bias=not batch_norm,
kernel_initializer=init_normal(),
name='conv'),
name='bn'))
csize *= 2
nfilt //= 2
# extra layers
for it in range(self.n_extra_generator_layers):
with tf.variable_scope('extra-{0}.{1}'.format(it, nfilt)):
tensor = tf.nn.relu(
bn(tf.layers.conv2d_transpose(
tensor,
nfilt,
3,
1,
'same',
use_bias=not batch_norm,
kernel_initializer=init_normal(),
name='conv'),
name='bn'))
# final layer
with tf.variable_scope('final.{0}-{1}'.format(
nfilt, self.n_channel)):
tensor = tf.layers.conv2d_transpose(
tensor,
self.n_channel,
4,
2,
'same',
activation=tf.tanh,
kernel_initializer=init_normal(),
name='conv')
# mask layer
return tensor
def _build_discriminator_base(self,
tensor=None,
training=False,
batch_norm=None):
nfilt = 500
if tensor is None:
tensor = self.input_x
if batch_norm is None:
batch_norm = self.batch_norm_D
if batch_norm:
def bn(tensor, name=None):
return tf.contrib.layers.batch_norm(tensor,
is_training=training,
renorm=BATCH_RENORM,
decay=BATCH_NORM_DECAY)
else:
bn = tf.identity
# initial layer
with tf.variable_scope('initial.{0}-{1}'.format(self.n_channel,
nfilt)):
tensor = lrelu(
bn(tf.layers.conv2d(tensor,
nfilt,
4,
2,
'same',
use_bias=not batch_norm,
kernel_initializer=init_normal(),
name='conv'),
name='bn'))
nfilt /= 2
csize = self.n_pixel / 2
# extra layers
for it in range(self.n_extra_discriminator_layers):
with tf.variable_scope('extra-{0}.{1}'.format(it, nfilt)):
tensor = lrelu(
bn(tf.layers.conv2d(tensor,
nfilt,
3,
1,
'same',
use_bias=not batch_norm,
kernel_initializer=init_normal(),
name='conv'),
name='bn'))
# downscaling layers
while csize > 4:
with tf.variable_scope('pyramid.{0}-{1}'.format(nfilt, nfilt * 2)):
tensor = lrelu(
bn(tf.layers.conv2d(tensor,
nfilt * 2,
4,
2,
'same',
use_bias=not batch_norm,
kernel_initializer=init_normal(),
name='conv'),
name='bn'))
nfilt *= 2
csize /= 2
return tensor
def _build_discriminator(self,
tensor=None,
label=None,
training=False,
batch_norm=None):
with tf.variable_scope('discriminator') as scope:
# set reuse if necessary
if tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES,
scope=scope.name):
scope.reuse_variables()
nfilt = 512
# this means that labeled real data is inputted
if tensor is None and label is None:
# add labels to input
x_shape = self.input_labeled_x.get_shape()
# WHY??
#tensor = tf.concat([self.input_labeled_x,
# tf.reshape(self.input_labeled_y, [-1, 1, 1, self.n_class]) * tf.ones(
# [x_shape[0], x_shape[1], x_shape[2], self.n_class])], axis=3)
tensor = tf.concat([
self.input_labeled_x,
tf.tile(
tf.reshape(self.input_labeled_y,
[-1, 1, 1, self.n_class]),
[1, x_shape[1], x_shape[2], 1])
],
axis=3)
elif tensor is None or label is None:
print('Tensor and label must be both None or both exists')
raise
else:
x_shape = tensor.get_shape()
print(label.get_shape())
label_copy = tf.tile(
tf.reshape(label, [-1, 1, 1, self.n_class]),
[1, x_shape[1], x_shape[2], 1])
tensor = tf.concat([tensor, label_copy], axis=3)
if batch_norm is None:
batch_norm = self.batch_norm_D
if batch_norm:
def bn(tensor, name=None):
return tf.contrib.layers.batch_norm(tensor,
is_training=training,
renorm=BATCH_RENORM,
decay=BATCH_NORM_DECAY)
else:
bn = tf.identity
# initial layer
with tf.variable_scope('initial.{0}-{1}'.format(
self.n_channel + self.n_class, nfilt)):
tensor = lrelu(
bn(tf.layers.conv2d(tensor,
nfilt,
4,
2,
'same',
use_bias=not batch_norm,
kernel_initializer=init_normal(),
name='conv'),
name='bn'))
nfilt //= 2
csize = self.n_pixel // 2
# extra layers
for it in range(self.n_extra_discriminator_layers):
with tf.variable_scope('extra-{0}.{1}'.format(it, nfilt)):
tensor = lrelu(
bn(tf.layers.conv2d(tensor,
nfilt,
3,
1,
'same',
use_bias=not batch_norm,
kernel_initializer=init_normal(),
name='conv'),
name='bn'))
# downscaling layers
while csize > 4:
with tf.variable_scope('pyramid.{0}-{1}'.format(
nfilt, nfilt * 2)):
tensor = lrelu(
bn(tf.layers.conv2d(tensor,
nfilt * 2,
4,
2,
'same',
use_bias=not batch_norm,
kernel_initializer=init_normal(),
name='conv'),
name='bn'))
nfilt *= 2
csize /= 2
# final layer
d_out = 2
with tf.variable_scope('final.{0}-{1}'.format(
tensor.shape[-1], d_out)):
out_logits = tf.reshape(
tf.reduce_mean(tf.layers.conv2d(
tensor,
d_out,
4,
2,
'valid',
kernel_initializer=init_normal(),
name='conv'),
axis=3), [-1, d_out])
# TODO consider to output the labels instead of softmax&out_logits
return tf.nn.softmax(out_logits), out_logits
def _build_generator(self,
tensor=None,
label=None,
training=False,
batch_norm=None):
assert self.n_pixel % 16 == 0, "isize has to be a multiple of 16"
nfilt = 2048
csize = 4
if label is None:
if self.input_y_g is None:
#TO BE EDITED
#sample generated images
batch_size = (self.input_z_g).shape[0]
label = get_one_hot(np.repeat(np.tile(self.n_class),
(batch_size / self.m_class) + 1),
depth=self.n_class)
label = label[0:batch_size, :]
else:
#get label from input
label = self.input_y_g
if tensor is None:
# add label to noise
tensor = tf.concat([self.input_z_g, label], 1)
else:
# assuming tensor is a specific noise
tensor = tf.concat([tensor, label], 1)
#tensor = tf.concat([tensor, tf.one_hot(label, self.n_class)], 1)
if batch_norm is None:
batch_norm = self.batch_norm_G
if batch_norm:
def bn(x, name=None):
return tf.contrib.layers.batch_norm(x,
is_training=training,
renorm=BATCH_RENORM,
decay=BATCH_NORM_DECAY)
else:
# return the same if bn is not aactivated
bn = tf.identity
with tf.variable_scope('generator') as scope:
# set reuse if necessary
if tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES,
scope=scope.name):
scope.reuse_variables()
# initial layer
with tf.variable_scope('initial.{0}-{1}'.format(
self.n_noise + self.n_class, nfilt)):
tensor = tf.nn.relu(
bn(tf.layers.conv2d_transpose(
tf.reshape(tensor,
[-1, 1, 1, self.n_noise + self.n_class]),
nfilt,
4,
2,
'valid',
use_bias=not batch_norm,
kernel_initializer=init_normal(),
name='conv'),
name='bn'))
# upscaling layers
while csize < self.n_pixel / 2:
with tf.variable_scope('pyramid.{0}-{1}'.format(
nfilt, nfilt // 2)):
tensor = tf.nn.relu(
bn(tf.layers.conv2d_transpose(
tensor,
nfilt // 2,
4,
2,
'same',
use_bias=not batch_norm,
kernel_initializer=init_normal(),
name='conv'),
name='bn'))
csize *= 2
nfilt //= 2
# extra layers
for it in range(self.n_extra_generator_layers):
with tf.variable_scope('extra-{0}.{1}'.format(it, nfilt)):
tensor = tf.nn.relu(
bn(tf.layers.conv2d_transpose(
tensor,
nfilt,
3,
1,
'same',
use_bias=not batch_norm,
kernel_initializer=init_normal(),
name='conv'),
name='bn'))
# TODO in original DCGAN struct, result is flattenned(3->1), not setting filtercount to 1 (self.n_channel)
# final layer
with tf.variable_scope('final.{0}-{1}'.format(
nfilt, self.n_channel)):
tensor = tf.layers.conv2d_transpose(
tensor,
self.n_channel,
4,
2,
'same',
activation=tf.tanh,
kernel_initializer=init_normal(),
name='conv')
# mask layer
return tensor * self.mask, label
def _build_classifier_base(self,
tensor=None,
training=False,
batch_norm=None):
nfilt = 128
if tensor is None:
tensor = self.input_x_c
if batch_norm is None:
batch_norm = self.batch_norm_D
if batch_norm:
def bn(tensor, name=None):
return tf.contrib.layers.batch_norm(tensor,
is_training=training,
renorm=BATCH_RENORM,
decay=BATCH_NORM_DECAY)
else:
bn = tf.identity
# tf.layers.conv2d(inputs,filters,kernel_size,strides,padding...)
def do(tensor, rate=0.5, name=None):
return tf.contrib.layers.dropout(tensor,
keep_prob=rate,
is_training=training)
# dropout before layers
with tf.variable_scope('initial_dropout{0}-{1}'.format(
self.n_channel, nfilt)):
tensor = tf.layers.dropout(tensor,
rate=0.2,
training=training,
seed=self.seed,
name='do')
# initial layer
for it in range(2):
with tf.variable_scope('first_part-{0}.{1}-{2}'.format(
it, self.n_channel, nfilt)):
tensor = lrelu(
bn(tf.layers.conv2d(tensor,
nfilt,
3,
1,
'same',
use_bias=not batch_norm,
kernel_initializer=init_normal(),
name='conv'),
name='bn'))
with tf.variable_scope('first_part-last{0}-{1}'.format(
self.n_channel, nfilt)):
tensor = do(lrelu(
bn(tf.layers.conv2d(tensor,
nfilt,
3,
2,
'same',
use_bias=not batch_norm,
kernel_initializer=init_normal(),
name='conv'),
name='bn')),
rate=0.5,
name='do')
nfilt = 256
for it in range(2):
with tf.variable_scope('second_part-{0}.{1}-{2}'.format(
it, self.n_channel, nfilt)):
tensor = lrelu(
bn(tf.layers.conv2d(tensor,
nfilt,
3,
1,
'same',
use_bias=not batch_norm,
kernel_initializer=init_normal(),
name='conv'),
name='bn'))
with tf.variable_scope('second_part-last{0}-{1}'.format(
self.n_channel, nfilt)):
tensor = do(lrelu(
bn(tf.layers.conv2d(tensor,
nfilt,
3,
2,
'same',
use_bias=not batch_norm,
kernel_initializer=init_normal(),
name='conv'),
name='bn')),
rate=0.5,
name='do')
nfilt = 512
with tf.variable_scope('third_part{0}-{1}'.format(
self.n_channel, nfilt)):
tensor = lrelu(
bn(tf.layers.conv2d(tensor,
nfilt,
3,
1,
'same',
use_bias=not batch_norm,
kernel_initializer=init_normal(),
name='conv'),
name='bn'))
nfilt = 256
with tf.variable_scope('third_part{0}-{1}'.format(
self.n_channel, nfilt)):
tensor = lrelu(
bn(tf.layers.conv2d(tensor,
nfilt,
1,
1,
'same',
use_bias=not batch_norm,
kernel_initializer=init_normal(),
name='conv'),
name='bn'))
nfilt = 128
with tf.variable_scope('third_part{0}-{1}'.format(
self.n_channel, nfilt)):
tensor = lrelu(
bn(tf.layers.conv2d(tensor,
nfilt,
1,
1,
'same',
use_bias=not batch_norm,
kernel_initializer=init_normal(),
name='conv'),
name='bn'))
with tf.variable_scope('last_layer{0}-{1}'.format(
self.n_channel, nfilt)):
tensor = tf.reduce_mean(tensor, [1, 2], name='rm')
tensor = lrelu(tf.layers.dense(tensor, self.n_class))
# csize = self.n_pixel / 2
# extra layers
# for it in range(self.n_extra_discriminator_layers):
# with tf.variable_scope('extra-{0}.{1}'.format(it, nfilt)):
# tensor = lrelu(bn(tf.layers.conv2d(tensor, nfilt, 3, 1, 'same', use_bias=not batch_norm,
# kernel_initializer=init_normal(),
# name='conv'), name='bn'))
# downscaling layers
# while csize > 4:
# with tf.variable_scope('pyramid.{0}-{1}'.format(nfilt, nfilt * 2)):
# tensor = lrelu(bn(tf.layers.conv2d(tensor, nfilt * 2, 4, 2, 'same', use_bias=not batch_norm,
# kernel_initializer=init_normal(),
# name='conv'), name='bn'))
# nfilt *= 2
# csize /= 2
return tensor