def decoder(self, z, name='decoder', is_reuse=False):
with tf.variable_scope(name) as scope:
if is_reuse is True:
scope.reuse_variables()
tf_utils.print_activations(z)
# 1st hidden layer
h0_linear = tf_utils.linear(z, self.n_hidden, name='h0_linear')
h0_tanh = tf_utils.tanh(h0_linear, name='h0_tanh')
h0_drop = tf.nn.dropout(h0_tanh,
keep_prob=self.keep_prob_tfph,
name='h0_drop')
tf_utils.print_activations(h0_drop)
# 2nd hidden layer
h1_linear = tf_utils.linear(h0_drop,
self.n_hidden,
name='h1_linear')
h1_elu = tf_utils.elu(h1_linear, name='h1_elu')
h1_drop = tf.nn.dropout(h1_elu,
keep_prob=self.keep_prob_tfph,
name='h1_drop')
tf_utils.print_activations(h1_drop)
# 3rd hidden layer
h2_linear = tf_utils.linear(h1_drop,
self.output_dim,
name='h2_linear')
h2_sigmoid = tf_utils.sigmoid(h2_linear, name='h2_sigmoid')
tf_utils.print_activations(h2_sigmoid)
output = tf.reshape(h2_sigmoid, [-1, *self.image_size])
tf_utils.print_activations(output)
return output
def basicGenerator(self, data, name='g_'):
with tf.variable_scope(name):
data_flatten = flatten(data)
tf_utils.print_activations(data_flatten)
# from (N, 128) to (N, 4, 4, 256)
h0_linear = tf_utils.linear(data_flatten,
self.gen_c[0],
name='h0_linear')
if self.flags.dataset == 'cifar10':
h0_linear = tf.reshape(h0_linear, [
tf.shape(h0_linear)[0], 4, 4,
int(self.gen_c[0] / (4 * 4))
])
h0_linear = tf_utils.norm(h0_linear,
_type='batch',
_ops=self.gen_train_ops,
name='h0_norm')
h0_relu = tf.nn.relu(h0_linear, name='h0_relu')
h0_reshape = tf.reshape(
h0_relu,
[tf.shape(h0_relu)[0], 4, 4,
int(self.gen_c[0] / (4 * 4))])
# from (N, 4, 4, 256) to (N, 8, 8, 128)
h1_deconv = tf_utils.deconv2d(h0_reshape,
self.gen_c[1],
k_h=5,
k_w=5,
name='h1_deconv2d')
if self.flags.dataset == 'cifar10':
h1_deconv = tf_utils.norm(h1_deconv,
_type='batch',
_ops=self.gen_train_ops,
name='h1_norm')
h1_relu = tf.nn.relu(h1_deconv, name='h1_relu')
# from (N, 8, 8, 128) to (N, 16, 16, 64)
h2_deconv = tf_utils.deconv2d(h1_relu,
self.gen_c[2],
k_h=5,
k_w=5,
name='h2_deconv2d')
if self.flags.dataset == 'cifar10':
h2_deconv = tf_utils.norm(h2_deconv,
_type='batch',
_ops=self.gen_train_ops,
name='h2_norm')
h2_relu = tf.nn.relu(h2_deconv, name='h2_relu')
# from (N, 16, 16, 64) to (N, 32, 32, 1)
output = tf_utils.deconv2d(h2_relu,
self.image_size[2],
k_h=5,
k_w=5,
name='h3_deconv2d')
return tf_utils.tanh(output)
def __call__(self, x):
with tf.variable_scope(self.name, reuse=self.reuse):
tf_utils.print_activations(x)
# (N, H, W, C) -> (N, H, W, 64)
conv1 = tf_utils.padding2d(x, p_h=3, p_w=3, pad_type='REFLECT', name='conv1_padding')
conv1 = tf_utils.conv2d(conv1, self.ngf, k_h=7, k_w=7, d_h=1, d_w=1, padding='VALID',
name='conv1_conv')
conv1 = tf_utils.norm(conv1, _type='instance', _ops=self._ops, name='conv1_norm')
conv1 = tf_utils.relu(conv1, name='conv1_relu', is_print=True)
# (N, H, W, 64) -> (N, H/2, W/2, 128)
conv2 = tf_utils.conv2d(conv1, 2*self.ngf, k_h=3, k_w=3, d_h=2, d_w=2, padding='SAME',
name='conv2_conv')
conv2 = tf_utils.norm(conv2, _type='instance', _ops=self._ops, name='conv2_norm',)
conv2 = tf_utils.relu(conv2, name='conv2_relu', is_print=True)
# (N, H/2, W/2, 128) -> (N, H/4, W/4, 256)
conv3 = tf_utils.conv2d(conv2, 4*self.ngf, k_h=3, k_w=3, d_h=2, d_w=2, padding='SAME',
name='conv3_conv')
conv3 = tf_utils.norm(conv3, _type='instance', _ops=self._ops, name='conv3_norm',)
conv3 = tf_utils.relu(conv3, name='conv3_relu', is_print=True)
# (N, H/4, W/4, 256) -> (N, H/4, W/4, 256)
if (self.image_size[0] <= 128) and (self.image_size[1] <= 128):
# use 6 residual blocks for 128x128 images
res_out = tf_utils.n_res_blocks(conv3, num_blocks=6, is_print=True)
else:
# use 9 blocks for higher resolution
res_out = tf_utils.n_res_blocks(conv3, num_blocks=9, is_print=True)
# (N, H/4, W/4, 256) -> (N, H/2, W/2, 128)
conv4 = tf_utils.deconv2d(res_out, 2*self.ngf, name='conv4_deconv2d')
conv4 = tf_utils.norm(conv4, _type='instance', _ops=self._ops, name='conv4_norm')
conv4 = tf_utils.relu(conv4, name='conv4_relu', is_print=True)
# (N, H/2, W/2, 128) -> (N, H, W, 64)
conv5 = tf_utils.deconv2d(conv4, self.ngf, name='conv5_deconv2d')
conv5 = tf_utils.norm(conv5, _type='instance', _ops=self._ops, name='conv5_norm')
conv5 = tf_utils.relu(conv5, name='conv5_relu', is_print=True)
# (N, H, W, 64) -> (N, H, W, 3)
conv6 = tf_utils.padding2d(conv5, p_h=3, p_w=3, pad_type='REFLECT', name='output_padding')
conv6 = tf_utils.conv2d(conv6, self.image_size[2], k_h=7, k_w=7, d_h=1, d_w=1,
padding='VALID', name='output_conv')
output = tf_utils.tanh(conv6, name='output_tanh', is_print=True)
# set reuse=True for next call
self.reuse = True
self.variables = tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES, scope=self.name)
return output
def encoder(self, data, name='encoder'):
with tf.variable_scope(name):
data_flatten = flatten(data)
tf_utils.print_activations(data_flatten)
# 1st hidden layer
h0_linear = tf_utils.linear(data_flatten,
self.n_hidden,
name='h0_linear')
h0_elu = tf_utils.elu(h0_linear, name='h0_elu')
h0_drop = tf.nn.dropout(h0_elu,
keep_prob=self.keep_prob_tfph,
name='h0_drop')
tf_utils.print_activations(h0_drop)
# 2nd hidden layer
h1_linear = tf_utils.linear(h0_drop,
self.n_hidden,
name='h1_linear')
h1_tanh = tf_utils.tanh(h1_linear, name='h1_tanh')
h1_drop = tf.nn.dropout(h1_tanh,
keep_prob=self.keep_prob_tfph,
name='h1_drop')
tf_utils.print_activations(h1_drop)
# 3rd hidden layer
h2_linear = tf_utils.linear(h1_drop,
2 * self.flags.z_dim,
name='h2_linear')
tf_utils.print_activations(h2_linear)
# The mean parameter is unconstrained
mean = h2_linear[:, :self.flags.z_dim]
# The standard deviation must be positive.
# Parameterize with a softplus and add a small epsilon for numerical stability
stddev = 1e-6 + tf.nn.softplus(h2_linear[:, self.flags.z_dim:])
tf_utils.print_activations(mean)
tf_utils.print_activations(stddev)
return mean, stddev
def __call__(self, i_to_s, state, name='DiagonalBiLSTMCell'):
c_prev = tf.slice(state, begin=[0, 0], size=[-1, self._num_units])
# [batch, height * hidden_dims]
h_prev = tf.slice(state,
begin=[0, self._num_units],
size=[-1, self._num_units])
# i_to_s: [batch, 4 * height * hidden_dims]
input_size = i_to_s.get_shape().with_rank(2)[1]
if input_size.value is None:
raise ValueError(
"Could not infer input size from inputs.get_shape()[-1]")
with tf.variable_scope(name):
# input-to-state (K_ss * h_{i-1}) : 2x1 convolution. generate 4h x n x n ternsor.
# [batch, height, 1, hidden_dims]
conv1d_inputs = tf.reshape(
h_prev, [-1, self._height, 1, self._hidden_dims],
name='conv1d_inputs')
# [batch, height, 1, hidden_dims * 4]
conv_s_to_s = tf_utils.conv1d(conv1d_inputs,
4 * self._hidden_dims,
kernel_size=2,
name='s_to_s')
# [batch, height * hidden_dims * 4]
s_to_s = tf.reshape(conv_s_to_s,
[-1, self._height * self._hidden_dims * 4])
lstm_matrix = tf_utils.sigmoid(s_to_s + i_to_s)
# i=input_gate, g=new_input, f=forget_gate, o=output_gate
o, f, i, g = tf.split(lstm_matrix, 4, axis=1)
c = f * c_prev + i * g
h = o * tf_utils.tanh(c)
new_state = tf.concat([c, h], axis=1)
return h, new_state
def __call__(self, x):
with tf.variable_scope(self.name, reuse=self.reuse):
tf_utils.print_activations(x)
# conv: (N, H, W, C) -> (N, H/2, W/2, 64)
output = tf_utils.conv2d(x,
self.conv_dims[0],
k_h=4,
k_w=4,
d_h=2,
d_w=2,
padding='SAME',
name='conv0_conv2d')
output = tf_utils.lrelu(output, name='conv0_lrelu', is_print=True)
for idx, conv_dim in enumerate(self.conv_dims[1:]):
# conv: (N, H/2, W/2, C) -> (N, H/4, W/4, 2C)
output = tf_utils.conv2d(output,
conv_dim,
k_h=4,
k_w=4,
d_h=2,
d_w=2,
padding='SAME',
name='conv{}_conv2d'.format(idx + 1))
output = tf_utils.norm(output,
_type=self.norm,
_ops=self._ops,
name='conv{}_norm'.format(idx + 1))
output = tf_utils.lrelu(output,
name='conv{}_lrelu'.format(idx + 1),
is_print=True)
for idx, deconv_dim in enumerate(self.deconv_dims):
# deconv: (N, H/16, W/16, C) -> (N, W/8, H/8, C/2)
output = tf_utils.deconv2d(output,
deconv_dim,
k_h=4,
k_w=4,
name='deconv{}_conv2d'.format(idx))
output = tf_utils.norm(output,
_type=self.norm,
_ops=self._ops,
name='deconv{}_norm'.format(idx))
output = tf_utils.relu(output,
name='deconv{}_relu'.format(idx),
is_print=True)
# conv: (N, H/2, W/2, 64) -> (N, W, H, 3)
output = tf_utils.deconv2d(output,
self.output_channel,
k_h=4,
k_w=4,
name='conv3_deconv2d')
output = tf_utils.tanh(output, name='conv4_tanh', is_print=True)
# set reuse=True for next call
self.reuse = True
self.variables = tf.get_collection(
tf.GraphKeys.TRAINABLE_VARIABLES, scope=self.name)
return output
def generator(self, data, name='g_'):
with tf.variable_scope(name):
data_flatten = flatten(data)
tf_utils.print_activations(data_flatten)
# from (N, 128) to (N, 2, 4, 512)
h0_linear = tf_utils.linear(data_flatten,
self.gen_c[0],
name='h0_linear')
h0_reshape = tf.reshape(
h0_linear,
[tf.shape(h0_linear)[0], 2, 4,
int(self.gen_c[0] / (2 * 4))])
# (N, 4, 8, 512)
resblock_1 = tf_utils.res_block_v2(h0_reshape,
self.gen_c[1],
filter_size=3,
_ops=self.gen_train_ops,
norm_='batch',
resample='up',
name='res_block_1')
# (N, 8, 16, 256)
resblock_2 = tf_utils.res_block_v2(resblock_1,
self.gen_c[2],
filter_size=3,
_ops=self.gen_train_ops,
norm_='batch',
resample='up',
name='res_block_2')
# (N, 16, 32, 128)
resblock_3 = tf_utils.res_block_v2(resblock_2,
self.gen_c[3],
filter_size=3,
_ops=self.gen_train_ops,
norm_='batch',
resample='up',
name='res_block_3')
# (N, 32, 64, 64)
resblock_4 = tf_utils.res_block_v2(resblock_3,
self.gen_c[4],
filter_size=3,
_ops=self.gen_train_ops,
norm_='batch',
resample='up',
name='res_block_4')
# (N, 64, 128, 32)
resblock_5 = tf_utils.res_block_v2(resblock_4,
self.gen_c[5],
filter_size=3,
_ops=self.gen_train_ops,
norm_='batch',
resample='up',
name='res_block_5')
# (N, 128, 256, 32)
resblock_6 = tf_utils.res_block_v2(resblock_5,
self.gen_c[6],
filter_size=3,
_ops=self.gen_train_ops,
norm_='batch',
resample='up',
name='res_block_6')
norm_7 = tf_utils.norm(resblock_6,
_type='batch',
_ops=self.gen_train_ops,
name='norm_7')
relu_7 = tf_utils.relu(norm_7, name='relu_7')
# (N, 128, 256, 3)
output = tf_utils.conv2d(relu_7,
output_dim=self.image_size[2],
k_w=3,
k_h=3,
d_h=1,
d_w=1,
name='output')
return tf_utils.tanh(output)
def __call__(self, x, keep_rate=0.5):
with tf.compat.v1.variable_scope(self.name, reuse=self.reuse):
tf_utils.print_activations(x, logger=self.logger)
# E0: (320, 200) -> (160, 100)
e0_conv2d = tf_utils.conv2d(x,
output_dim=self.gen_c[0],
initializer='He',
logger=self.logger,
name='e0_conv2d')
e0_lrelu = tf_utils.lrelu(e0_conv2d,
logger=self.logger,
name='e0_lrelu')
# E1: (160, 100) -> (80, 50)
e1_conv2d = tf_utils.conv2d(e0_lrelu,
output_dim=self.gen_c[1],
initializer='He',
logger=self.logger,
name='e1_conv2d')
e1_batchnorm = tf_utils.norm(e1_conv2d,
_type=self.norm,
_ops=self._ops,
logger=self.logger,
name='e1_norm')
e1_lrelu = tf_utils.lrelu(e1_batchnorm,
logger=self.logger,
name='e1_lrelu')
# E2: (80, 50) -> (40, 25)
e2_conv2d = tf_utils.conv2d(e1_lrelu,
output_dim=self.gen_c[2],
initializer='He',
logger=self.logger,
name='e2_conv2d')
e2_batchnorm = tf_utils.norm(e2_conv2d,
_type=self.norm,
_ops=self._ops,
logger=self.logger,
name='e2_norm')
e2_lrelu = tf_utils.lrelu(e2_batchnorm,
logger=self.logger,
name='e2_lrelu')
# E3: (40, 25) -> (20, 13)
e3_conv2d = tf_utils.conv2d(e2_lrelu,
output_dim=self.gen_c[3],
initializer='He',
logger=self.logger,
name='e3_conv2d')
e3_batchnorm = tf_utils.norm(e3_conv2d,
_type=self.norm,
_ops=self._ops,
logger=self.logger,
name='e3_norm')
e3_lrelu = tf_utils.lrelu(e3_batchnorm,
logger=self.logger,
name='e3_lrelu')
# E4: (20, 13) -> (10, 7)
e4_conv2d = tf_utils.conv2d(e3_lrelu,
output_dim=self.gen_c[4],
initializer='He',
logger=self.logger,
name='e4_conv2d')
e4_batchnorm = tf_utils.norm(e4_conv2d,
_type=self.norm,
_ops=self._ops,
logger=self.logger,
name='e4_norm')
e4_lrelu = tf_utils.lrelu(e4_batchnorm,
logger=self.logger,
name='e4_lrelu')
# E5: (10, 7) -> (5, 4)
e5_conv2d = tf_utils.conv2d(e4_lrelu,
output_dim=self.gen_c[5],
initializer='He',
logger=self.logger,
name='e5_conv2d')
e5_batchnorm = tf_utils.norm(e5_conv2d,
_type=self.norm,
_ops=self._ops,
logger=self.logger,
name='e5_norm')
e5_lrelu = tf_utils.lrelu(e5_batchnorm,
logger=self.logger,
name='e5_lrelu')
# E6: (5, 4) -> (3, 2)
e6_conv2d = tf_utils.conv2d(e5_lrelu,
output_dim=self.gen_c[6],
initializer='He',
logger=self.logger,
name='e6_conv2d')
e6_batchnorm = tf_utils.norm(e6_conv2d,
_type=self.norm,
_ops=self._ops,
logger=self.logger,
name='e6_norm')
e6_lrelu = tf_utils.lrelu(e6_batchnorm,
logger=self.logger,
name='e6_lrelu')
# E7: (3, 2) -> (2, 1)
e7_conv2d = tf_utils.conv2d(e6_lrelu,
output_dim=self.gen_c[7],
initializer='He',
logger=self.logger,
name='e7_conv2d')
e7_batchnorm = tf_utils.norm(e7_conv2d,
_type=self.norm,
_ops=self._ops,
logger=self.logger,
name='e7_norm')
e7_relu = tf_utils.lrelu(e7_batchnorm,
logger=self.logger,
name='e7_relu')
# D0: (2, 1) -> (3, 2)
# Stage1: (2, 1) -> (4, 2)
d0_deconv = tf_utils.deconv2d(e7_relu,
output_dim=self.gen_c[8],
initializer='He',
logger=self.logger,
name='d0_deconv2d')
# Stage2: (4, 2) -> (3, 2)
shapeA = e6_conv2d.get_shape().as_list()[1]
shapeB = d0_deconv.get_shape().as_list()[1] - e6_conv2d.get_shape(
).as_list()[1]
d0_split, _ = tf.split(d0_deconv, [shapeA, shapeB],
axis=1,
name='d0_split')
tf_utils.print_activations(d0_split, logger=self.logger)
# Stage3: Batch norm, concatenation, and relu
d0_batchnorm = tf_utils.norm(d0_split,
_type=self.norm,
_ops=self._ops,
logger=self.logger,
name='d0_norm')
d0_drop = tf_utils.dropout(d0_batchnorm,
keep_prob=keep_rate,
logger=self.logger,
name='d0_dropout')
d0_concat = tf.concat([d0_drop, e6_batchnorm],
axis=3,
name='d0_concat')
d0_relu = tf_utils.relu(d0_concat,
logger=self.logger,
name='d0_relu')
# D1: (3, 2) -> (5, 4)
# Stage1: (3, 2) -> (6, 4)
d1_deconv = tf_utils.deconv2d(d0_relu,
output_dim=self.gen_c[9],
initializer='He',
logger=self.logger,
name='d1_deconv2d')
# Stage2: (6, 4) -> (5, 4)
shapeA = e5_batchnorm.get_shape().as_list()[1]
shapeB = d1_deconv.get_shape().as_list(
)[1] - e5_batchnorm.get_shape().as_list()[1]
d1_split, _ = tf.split(d1_deconv, [shapeA, shapeB],
axis=1,
name='d1_split')
tf_utils.print_activations(d1_split, logger=self.logger)
# Stage3: Batch norm, concatenation, and relu
d1_batchnorm = tf_utils.norm(d1_split,
_type=self.norm,
_ops=self._ops,
logger=self.logger,
name='d1_norm')
d1_drop = tf_utils.dropout(d1_batchnorm,
keep_prob=keep_rate,
logger=self.logger,
name='d1_dropout')
d1_concat = tf.concat([d1_drop, e5_batchnorm],
axis=3,
name='d1_concat')
d1_relu = tf_utils.relu(d1_concat,
logger=self.logger,
name='d1_relu')
# D2: (5, 4) -> (10, 7)
# Stage1: (5, 4) -> (10, 8)
d2_deconv = tf_utils.deconv2d(d1_relu,
output_dim=self.gen_c[10],
initializer='He',
logger=self.logger,
name='d2_deconv2d')
# Stage2: (10, 8) -> (10, 7)
shapeA = e4_batchnorm.get_shape().as_list()[2]
shapeB = d2_deconv.get_shape().as_list(
)[2] - e4_batchnorm.get_shape().as_list()[2]
d2_split, _ = tf.split(d2_deconv, [shapeA, shapeB],
axis=2,
name='d2_split')
tf_utils.print_activations(d2_split, logger=self.logger)
# Stage3: Batch norm, concatenation, and relu
d2_batchnorm = tf_utils.norm(d2_split,
_type=self.norm,
_ops=self._ops,
logger=self.logger,
name='d2_norm')
d2_drop = tf_utils.dropout(d2_batchnorm,
keep_prob=keep_rate,
logger=self.logger,
name='d2_dropout')
d2_concat = tf.concat([d2_drop, e4_batchnorm],
axis=3,
name='d2_concat')
d2_relu = tf_utils.relu(d2_concat,
logger=self.logger,
name='d2_relu')
# D3: (10, 7) -> (20, 13)
# Stage1: (10, 7) -> (20, 14)
d3_deconv = tf_utils.deconv2d(d2_relu,
output_dim=self.gen_c[11],
initializer='He',
logger=self.logger,
name='d3_deconv2d')
# Stage2: (20, 14) -> (20, 13)
shapeA = e3_batchnorm.get_shape().as_list()[2]
shapeB = d3_deconv.get_shape().as_list(
)[2] - e3_batchnorm.get_shape().as_list()[2]
d3_split, _ = tf.split(d3_deconv, [shapeA, shapeB],
axis=2,
name='d3_split_2')
tf_utils.print_activations(d3_split, logger=self.logger)
# Stage3: Batch norm, concatenation, and relu
d3_batchnorm = tf_utils.norm(d3_split,
_type=self.norm,
_ops=self._ops,
logger=self.logger,
name='d3_norm')
d3_concat = tf.concat([d3_batchnorm, e3_batchnorm],
axis=3,
name='d3_concat')
d3_relu = tf_utils.relu(d3_concat,
logger=self.logger,
name='d3_relu')
# D4: (20, 13) -> (40, 25)
# Stage1: (20, 13) -> (40, 26)
d4_deconv = tf_utils.deconv2d(d3_relu,
output_dim=self.gen_c[12],
initializer='He',
logger=self.logger,
name='d4_deconv2d')
# Stage2: (40, 26) -> (40, 25)
shapeA = e2_batchnorm.get_shape().as_list()[2]
shapeB = d4_deconv.get_shape().as_list(
)[2] - e2_batchnorm.get_shape().as_list()[2]
d4_split, _ = tf.split(d4_deconv, [shapeA, shapeB],
axis=2,
name='d4_split')
tf_utils.print_activations(d4_split, logger=self.logger)
# Stage3: Batch norm, concatenation, and relu
d4_batchnorm = tf_utils.norm(d4_split,
_type=self.norm,
_ops=self._ops,
logger=self.logger,
name='d4_norm')
d4_concat = tf.concat([d4_batchnorm, e2_batchnorm],
axis=3,
name='d4_concat')
d4_relu = tf_utils.relu(d4_concat,
logger=self.logger,
name='d4_relu')
# D5: (40, 25, 256) -> (80, 50, 128)
d5_deconv = tf_utils.deconv2d(d4_relu,
output_dim=self.gen_c[13],
initializer='He',
logger=self.logger,
name='d5_deconv2d')
d5_batchnorm = tf_utils.norm(d5_deconv,
_type=self.norm,
_ops=self._ops,
logger=self.logger,
name='d5_norm')
d5_concat = tf.concat([d5_batchnorm, e1_batchnorm],
axis=3,
name='d5_concat')
d5_relu = tf_utils.relu(d5_concat,
logger=self.logger,
name='d5_relu')
# D6: (80, 50, 128) -> (160, 100, 64)
d6_deconv = tf_utils.deconv2d(d5_relu,
output_dim=self.gen_c[14],
initializer='He',
logger=self.logger,
name='d6_deconv2d')
d6_batchnorm = tf_utils.norm(d6_deconv,
_type=self.norm,
_ops=self._ops,
logger=self.logger,
name='d6_norm')
d6_concat = tf.concat([d6_batchnorm, e0_conv2d],
axis=3,
name='d6_concat')
d6_relu = tf_utils.relu(d6_concat,
logger=self.logger,
name='d6_relu')
# D7: (160, 100, 64) -> (320, 200, 1)
d7_deconv = tf_utils.deconv2d(d6_relu,
output_dim=self.gen_c[15],
initializer='He',
logger=self.logger,
name='d7_deconv2d')
output = tf_utils.tanh(d7_deconv,
logger=self.logger,
name='output_tanh')
# Set reuse=True for next call
self.reuse = True
self.variables = tf.compat.v1.get_collection(
tf.compat.v1.GraphKeys.TRAINABLE_VARIABLES, scope=self.name)
return output
def __call__(self, x):
with tf.variable_scope(self.name, reuse=self.reuse):
tf_utils.print_activations(x)
# conv: (N, H, W, C) -> (N, H/2, W/2, 64)
output = tf_utils.conv2d(x,
self.conv_dims[0],
k_h=4,
k_w=4,
d_h=2,
d_w=2,
padding='SAME',
name='conv0_conv2d')
output = tf_utils.lrelu(output, name='conv0_lrelu', is_print=True)
for idx, conv_dim in enumerate(self.conv_dims[1:]):
# conv: (N, H/2, W/2, C) -> (N, H/4, W/4, 2C)
output = tf_utils.conv2d(output,
conv_dim,
k_h=4,
k_w=4,
d_h=2,
d_w=2,
padding='SAME',
name='conv{}_conv2d'.format(idx + 1))
output = tf_utils.norm(output,
_type=self.norm,
_ops=self._ops,
name='conv{}_norm'.format(idx + 1))
output = tf_utils.lrelu(output,
name='conv{}_lrelu'.format(idx + 1),
is_print=True)
for idx, deconv_dim in enumerate(self.deconv_dims):
# deconv: (N, H/16, W/16, C) -> (N, W/8, H/8, C/2)
output = tf_utils.deconv2d(output,
deconv_dim,
k_h=4,
k_w=4,
name='deconv{}_conv2d'.format(idx))
output = tf_utils.norm(output,
_type=self.norm,
_ops=self._ops,
name='deconv{}_norm'.format(idx))
output = tf_utils.relu(output,
name='deconv{}_relu'.format(idx),
is_print=True)
# split (N, 152, 104, 64) to (N, 150, 104, 64)
shapeA = int(self.img_size[0] / 2)
shapeB = output.get_shape().as_list()[1] - shapeA
output, _ = tf.split(output, [shapeA, shapeB],
axis=1,
name='split_0')
tf_utils.print_activations(output)
# split (N, 150, 104, 64) to (N, 150, 100, 64)
shapeA = int(self.img_size[1] / 2)
shapeB = output.get_shape().as_list()[2] - shapeA
output, _ = tf.split(output, [shapeA, shapeB],
axis=2,
name='split_1')
tf_utils.print_activations(output)
# conv: (N, H/2, W/2, 64) -> (N, W, H, 3)
output = tf_utils.deconv2d(output,
self.img_size[2],
k_h=4,
k_w=4,
name='conv3_deconv2d')
output = tf_utils.tanh(output, name='conv4_tanh', is_print=True)
# set reuse=True for next call
self.reuse = True
self.variables = tf.get_collection(
tf.GraphKeys.TRAINABLE_VARIABLES, scope=self.name)
return output
def __call__(self, x, is_train=True):
with tf.variable_scope(self.name, reuse=self.reuse):
tf_utils.print_activations(x)
# (N, 100) -> (N, 4, 5, 1024)
h0_linear = tf_utils.linear(
x,
4 * 5 * self.dims[0],
name='h0_linear',
initializer='He',
logger=self.logger if is_train is True else None)
h0_reshape = tf.reshape(
h0_linear, [tf.shape(h0_linear)[0], 4, 5, self.dims[0]])
h0_norm = tf_utils.norm(
h0_reshape,
name='h0_batch',
_type='batch',
_ops=self._ops,
is_train=is_train,
logger=self.logger if is_train is True else None)
h0_relu = tf_utils.relu(
h0_norm,
name='h0_relu',
logger=self.logger if is_train is True else None)
# (N, 4, 5, 1024) -> (N, 8, 10, 512)
h1_deconv = tf_utils.deconv2d(
h0_relu,
output_dim=self.dims[1],
name='h1_deconv2d',
initializer='He',
logger=self.logger if is_train is True else None)
h1_norm = tf_utils.norm(
h1_deconv,
name='h1_batch',
_type='batch',
_ops=self._ops,
is_train=is_train,
logger=self.logger if is_train is True else None)
h1_relu = tf_utils.relu(
h1_norm,
name='h1_relu',
logger=self.logger if is_train is True else None)
# (N, 8, 10, 512) -> (N, 16, 20, 256)
h2_deconv = tf_utils.deconv2d(
h1_relu,
output_dim=self.dims[2],
name='h2_deconv2d',
initializer='He',
logger=self.logger if is_train is True else None)
h2_norm = tf_utils.norm(
h2_deconv,
name='h2_batch',
_type='batch',
_ops=self._ops,
is_train=is_train,
logger=self.logger if is_train is True else None)
h2_relu = tf_utils.relu(
h2_norm,
name='h2_relu',
logger=self.logger if is_train is True else None)
# (N, 16, 20, 256) -> (N, 15, 20, 256)
h2_split, _ = tf.split(h2_relu, [15, 1], axis=1, name='h2_split')
tf_utils.print_activations(
h2_split, logger=self.logger if is_train is True else None)
# (N, 15, 20, 256) -> (N, 30, 40, 128)
h3_deconv = tf_utils.deconv2d(
h2_split,
output_dim=self.dims[3],
name='h3_deconv2d',
initializer='He',
logger=self.logger if is_train is True else None)
h3_norm = tf_utils.norm(
h3_deconv,
name='h3_batch',
_type='batch',
_ops=self._ops,
is_train=is_train,
logger=self.logger if is_train is True else None)
h3_relu = tf_utils.relu(
h3_norm,
name='h3_relu',
logger=self.logger if is_train is True else None)
# (N, 30, 40, 128) -> (N, 60, 80, 64)
h4_deconv = tf_utils.deconv2d(
h3_relu,
output_dim=self.dims[4],
name='h4_deconv2d',
initializer='He',
logger=self.logger if is_train is True else None)
h4_norm = tf_utils.norm(
h4_deconv,
name='h4_batch',
_type='batch',
_ops=self._ops,
is_train=is_train,
logger=self.logger if is_train is True else None)
h4_relu = tf_utils.relu(
h4_norm,
name='h4_relu',
logger=self.logger if is_train is True else None)
# (N, 60, 80, 64) -> (N, 120, 160, 1)
h5_deconv = tf_utils.deconv2d(
h4_relu,
output_dim=self.dims[5],
name='h5_deconv',
initializer='He',
logger=self.logger if is_train is True else None)
output = tf_utils.tanh(
h5_deconv,
name='output',
logger=self.logger if is_train is True else None)
# Set reuse=True for next call
self.reuse = True
self.variables = tf.get_collection(
tf.GraphKeys.TRAINABLE_VARIABLES, scope=self.name)
return output
def __call__(self, x, is_train=True):
with tf.variable_scope(self.name, reuse=self.reuse):
tf_utils.print_activations(x)
# (N, 100) -> (N, 4, 5, 512)
h0_linear = tf_utils.linear(
x,
4 * 5 * self.dims[0],
name='h0_linear',
initializer='He',
logger=self.logger if is_train is True else None)
h0_reshape = tf.reshape(
h0_linear, [tf.shape(h0_linear)[0], 4, 5, self.dims[0]])
# (N, 4, 5, 512) -> (N, 8, 10, 512)
resblock_1 = tf_utils.res_block_v2(
x=h0_reshape,
k=self.dims[1],
filter_size=3,
_ops=self._ops,
norm_='batch',
resample='up',
name='res_block_1',
logger=self.logger if is_train is True else None)
# (N, 8, 10, 512) -> (N, 16, 20, 256)
resblock_2 = tf_utils.res_block_v2(
x=resblock_1,
k=self.dims[2],
filter_size=3,
_ops=self._ops,
norm_='batch',
resample='up',
name='res_block_2',
logger=self.logger if is_train is True else None)
# (N, 16, 20, 256) -> (N, 15, 20, 256)
resblock_2_split, _ = tf.split(resblock_2, [15, 1],
axis=1,
name='resblock_2_split')
tf_utils.print_activations(
resblock_2_split,
logger=self.logger if is_train is True else None)
# (N, 15, 20, 256) -> (N, 30, 40, 128)
resblock_3 = tf_utils.res_block_v2(
x=resblock_2_split,
k=self.dims[3],
filter_size=3,
_ops=self._ops,
norm_='batch',
resample='up',
name='res_block_3',
logger=self.logger if is_train is True else None)
# (N, 30, 40, 128) -> (N, 60, 80, 64)
resblock_4 = tf_utils.res_block_v2(
x=resblock_3,
k=self.dims[4],
filter_size=3,
_ops=self._ops,
norm_='batch',
resample='up',
name='res_block_4',
logger=self.logger if is_train is True else None)
# (N, 60, 80, 64) -> (N, 120, 160, 64)
resblock_5 = tf_utils.res_block_v2(
x=resblock_4,
k=self.dims[5],
filter_size=3,
_ops=self._ops,
norm_='batch',
resample='up',
name='res_block_5',
logger=self.logger if is_train is True else None)
norm_5 = tf_utils.norm(
resblock_5,
name='norm_5',
_type='batch',
_ops=self._ops,
is_train=is_train,
logger=self.logger if is_train is True else None)
relu_5 = tf_utils.relu(
norm_5,
name='relu_5',
logger=self.logger if is_train is True else None)
# (N, 120, 160, 64) -> (N, 120, 160, 3)
conv_6 = tf_utils.conv2d(
relu_5,
output_dim=self.dims[6],
k_h=3,
k_w=3,
d_h=1,
d_w=1,
name='conv_6',
logger=self.logger if is_train is True else None)
output = tf_utils.tanh(
conv_6,
name='output',
logger=self.logger if is_train is True else None)
# Set reuse=True for next call
self.reuse = True
self.variables = tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES,
scope=self.name)
return output
def __call__(self, x):
with tf.variable_scope(self.name, reuse=self.reuse):
tf_utils.print_activations(x)
# (300, 200) -> (150, 100)
e0_conv2d = tf_utils.conv2d(x, self.gen_c[0], name='e0_conv2d')
e0_lrelu = tf_utils.lrelu(e0_conv2d, name='e0_lrelu')
# (150, 100) -> (75, 50)
e1_conv2d = tf_utils.conv2d(e0_lrelu,
self.gen_c[1],
name='e1_conv2d')
e1_batchnorm = tf_utils.batch_norm(e1_conv2d,
name='e1_batchnorm',
_ops=self._ops)
e1_lrelu = tf_utils.lrelu(e1_batchnorm, name='e1_lrelu')
# (75, 50) -> (38, 25)
e2_conv2d = tf_utils.conv2d(e1_lrelu,
self.gen_c[2],
name='e2_conv2d')
e2_batchnorm = tf_utils.batch_norm(e2_conv2d,
name='e2_batchnorm',
_ops=self._ops)
e2_lrelu = tf_utils.lrelu(e2_batchnorm, name='e2_lrelu')
# (38, 25) -> (19, 13)
e3_conv2d = tf_utils.conv2d(e2_lrelu,
self.gen_c[3],
name='e3_conv2d')
e3_batchnorm = tf_utils.batch_norm(e3_conv2d,
name='e3_batchnorm',
_ops=self._ops)
e3_lrelu = tf_utils.lrelu(e3_batchnorm, name='e3_lrelu')
# (19, 13) -> (10, 7)
e4_conv2d = tf_utils.conv2d(e3_lrelu,
self.gen_c[4],
name='e4_conv2d')
e4_batchnorm = tf_utils.batch_norm(e4_conv2d,
name='e4_batchnorm',
_ops=self._ops)
e4_lrelu = tf_utils.lrelu(e4_batchnorm, name='e4_lrelu')
# (10, 7) -> (5, 4)
e5_conv2d = tf_utils.conv2d(e4_lrelu,
self.gen_c[5],
name='e5_conv2d')
e5_batchnorm = tf_utils.batch_norm(e5_conv2d,
name='e5_batchnorm',
_ops=self._ops)
e5_lrelu = tf_utils.lrelu(e5_batchnorm, name='e5_lrelu')
# (5, 4) -> (3, 2)
e6_conv2d = tf_utils.conv2d(e5_lrelu,
self.gen_c[6],
name='e6_conv2d')
e6_batchnorm = tf_utils.batch_norm(e6_conv2d,
name='e6_batchnorm',
_ops=self._ops)
e6_lrelu = tf_utils.lrelu(e6_batchnorm, name='e6_lrelu')
# (3, 2) -> (2, 1)
e7_conv2d = tf_utils.conv2d(e6_lrelu,
self.gen_c[7],
name='e7_conv2d')
e7_batchnorm = tf_utils.batch_norm(e7_conv2d,
name='e7_batchnorm',
_ops=self._ops)
e7_relu = tf_utils.relu(e7_batchnorm, name='e7_relu')
# (2, 1) -> (4, 2)
d0_deconv = tf_utils.deconv2d(e7_relu,
self.gen_c[8],
name='d0_deconv2d')
shapeA = e6_conv2d.get_shape().as_list()[1]
shapeB = d0_deconv.get_shape().as_list()[1] - e6_conv2d.get_shape(
).as_list()[1]
# (4, 2) -> (3, 2)
d0_split, _ = tf.split(d0_deconv, [shapeA, shapeB],
axis=1,
name='d0_split')
tf_utils.print_activations(d0_split)
d0_batchnorm = tf_utils.batch_norm(d0_split,
name='d0_batchnorm',
_ops=self._ops)
d0_drop = tf.nn.dropout(d0_batchnorm,
keep_prob=0.5,
name='d0_dropout')
d0_concat = tf.concat([d0_drop, e6_batchnorm],
axis=3,
name='d0_concat')
d0_relu = tf_utils.relu(d0_concat, name='d0_relu')
# (3, 2) -> (6, 4)
d1_deconv = tf_utils.deconv2d(d0_relu,
self.gen_c[9],
name='d1_deconv2d')
# (6, 4) -> (5, 4)
shapeA = e5_batchnorm.get_shape().as_list()[1]
shapeB = d1_deconv.get_shape().as_list(
)[1] - e5_batchnorm.get_shape().as_list()[1]
d1_split, _ = tf.split(d1_deconv, [shapeA, shapeB],
axis=1,
name='d1_split')
tf_utils.print_activations(d1_split)
d1_batchnorm = tf_utils.batch_norm(d1_split,
name='d1_batchnorm',
_ops=self._ops)
d1_drop = tf.nn.dropout(d1_batchnorm,
keep_prob=0.5,
name='d1_dropout')
d1_concat = tf.concat([d1_drop, e5_batchnorm],
axis=3,
name='d1_concat')
d1_relu = tf_utils.relu(d1_concat, name='d1_relu')
# (5, 4) -> (10, 8)
d2_deconv = tf_utils.deconv2d(d1_relu,
self.gen_c[10],
name='d2_deconv2d')
# (10, 8) -> (10, 7)
shapeA = e4_batchnorm.get_shape().as_list()[2]
shapeB = d2_deconv.get_shape().as_list(
)[2] - e4_batchnorm.get_shape().as_list()[2]
d2_split, _ = tf.split(d2_deconv, [shapeA, shapeB],
axis=2,
name='d2_split')
tf_utils.print_activations(d2_split)
d2_batchnorm = tf_utils.batch_norm(d2_split,
name='d2_batchnorm',
_ops=self._ops)
d2_drop = tf.nn.dropout(d2_batchnorm,
keep_prob=0.5,
name='d2_dropout')
d2_concat = tf.concat([d2_drop, e4_batchnorm],
axis=3,
name='d2_concat')
d2_relu = tf_utils.relu(d2_concat, name='d2_relu')
# (10, 7) -> (20, 14)
d3_deconv = tf_utils.deconv2d(d2_relu,
self.gen_c[11],
name='d3_deconv2d')
# (20, 14) -> (19, 14)
shapeA = e3_batchnorm.get_shape().as_list()[1]
shapeB = d3_deconv.get_shape().as_list(
)[1] - e3_batchnorm.get_shape().as_list()[1]
d3_split_1, _ = tf.split(d3_deconv, [shapeA, shapeB],
axis=1,
name='d3_split_1')
tf_utils.print_activations(d3_split_1)
# (19, 14) -> (19, 13)
shapeA = e3_batchnorm.get_shape().as_list()[2]
shapeB = d3_split_1.get_shape().as_list(
)[2] - e3_batchnorm.get_shape().as_list()[2]
d3_split_2, _ = tf.split(d3_split_1, [shapeA, shapeB],
axis=2,
name='d3_split_2')
tf_utils.print_activations(d3_split_2)
d3_batchnorm = tf_utils.batch_norm(d3_split_2,
name='d3_batchnorm',
_ops=self._ops)
d3_concat = tf.concat([d3_batchnorm, e3_batchnorm],
axis=3,
name='d3_concat')
d3_relu = tf_utils.relu(d3_concat, name='d3_relu')
# (19, 13) -> (38, 26)
d4_deconv = tf_utils.deconv2d(d3_relu,
self.gen_c[12],
name='d4_deconv2d')
# (38, 26) -> (38, 25)
shapeA = e2_batchnorm.get_shape().as_list()[2]
shapeB = d4_deconv.get_shape().as_list(
)[2] - e2_batchnorm.get_shape().as_list()[2]
d4_split, _ = tf.split(d4_deconv, [shapeA, shapeB],
axis=2,
name='d4_split')
tf_utils.print_activations(d4_split)
d4_batchnorm = tf_utils.batch_norm(d4_split,
name='d4_batchnorm',
_ops=self._ops)
d4_concat = tf.concat([d4_batchnorm, e2_batchnorm],
axis=3,
name='d4_concat')
d4_relu = tf_utils.relu(d4_concat, name='d4_relu')
# (38, 25) -> (76, 50)
d5_deconv = tf_utils.deconv2d(d4_relu,
self.gen_c[13],
name='d5_deconv2d')
# (76, 50) -> (75, 50)
shapeA = e1_batchnorm.get_shape().as_list()[1]
shapeB = d5_deconv.get_shape().as_list(
)[1] - e1_batchnorm.get_shape().as_list()[1]
d5_split, _ = tf.split(d5_deconv, [shapeA, shapeB],
axis=1,
name='d5_split')
tf_utils.print_activations(d5_split)
d5_batchnorm = tf_utils.batch_norm(d5_split,
name='d5_batchnorm',
_ops=self._ops)
d5_concat = tf.concat([d5_batchnorm, e1_batchnorm],
axis=3,
name='d5_concat')
d5_relu = tf_utils.relu(d5_concat, name='d5_relu')
# (75, 50) -> (150, 100)
d6_deconv = tf_utils.deconv2d(d5_relu,
self.gen_c[14],
name='d6_deconv2d')
d6_batchnorm = tf_utils.batch_norm(d6_deconv,
name='d6_batchnorm',
_ops=self._ops)
d6_concat = tf.concat([d6_batchnorm, e0_conv2d],
axis=3,
name='d6_concat')
d6_relu = tf_utils.relu(d6_concat, name='d6_relu')
# (150, 100) -> (300, 200)
d7_deconv = tf_utils.deconv2d(d6_relu,
self.gen_c[15],
name='d7_deconv2d')
output = tf_utils.tanh(d7_deconv, name='output_tanh')
# set reuse=True for next call
self.reuse = True
self.variables = tf.get_collection(
tf.GraphKeys.TRAINABLE_VARIABLES, scope=self.name)
return output
