def bottleneck_block(self, inputs, filters, train_mode,
projection_shortcut, strides, name):
with tf.compat.v1.variable_scope(name):
shortcut = inputs
inputs = tf_utils.relu(inputs, name='relu_0', logger=None)
# The projection shortcut shouldcome after the first batch norm and ReLU since it perofrms a 1x1 convolution.
if projection_shortcut is not None:
shortcut = self.projection_shortcut(inputs=inputs,
filters_out=filters,
strides=strides,
name='conv_projection')
inputs = self.conv2d_fixed_padding(inputs=inputs,
filters=filters,
kernel_size=3,
strides=strides,
name='conv_0')
inputs = tf_utils.relu(inputs, name='relu_1', logger=None)
inputs = self.conv2d_fixed_padding(inputs=inputs,
filters=filters,
kernel_size=3,
strides=1,
name='conv_1')
output = tf.identity(inputs + shortcut, name=(name + '_output'))
tf_utils.print_activations(output, logger=None)
return output
def forward_network(self, input_img, reuse=False):
with tf.compat.v1.variable_scope(self.name, reuse=reuse):
tf_utils.print_activations(input_img, logger=None)
inputs = self.conv2d_fixed_padding(inputs=input_img, filters=64, kernel_size=7, strides=2, name='conv1')
inputs = tf_utils.max_pool(inputs, name='3x3_maxpool', ksize=[1, 3, 3, 1], strides=[1, 2, 2, 1],
logger=None)
inputs = self.block_layer(inputs=inputs, filters=64, block_fn=self.bottleneck_block, blocks=self.layers[0],
strides=1, train_mode=False, name='block_layer1')
inputs = self.block_layer(inputs=inputs, filters=128, block_fn=self.bottleneck_block, blocks=self.layers[1],
strides=2, train_mode=False, name='block_layer2')
inputs = self.block_layer(inputs=inputs, filters=256, block_fn=self.bottleneck_block, blocks=self.layers[2],
strides=2, train_mode=False, name='block_layer3')
inputs = self.block_layer(inputs=inputs, filters=512, block_fn=self.bottleneck_block, blocks=self.layers[3],
strides=2, train_mode=False, name='block_layer4')
inputs = tf_utils.relu(inputs, name='before_flatten_relu', logger=None)
# _, h, w, _ = inputs.get_shape().as_list()
# inputs = tf_utils.avg_pool(inputs, name='gap', ksize=[1, h, w, 1], strides=[1, 1, 1, 1], logger=self.logger)
# Flatten & FC1
inputs = tf_utils.flatten(inputs, name='flatten', logger=None)
inputs = tf_utils.linear(inputs, 512, name='FC1')
inputs = tf_utils.relu(inputs, name='FC1_relu', logger=None)
inputs = tf_utils.linear(inputs, 256, name='FC2')
inputs = tf_utils.relu(inputs, name='FC2_relu', logger=None)
logits = tf_utils.linear(inputs, self.num_attribute, name='Out')
return logits
def Generator_Softmax(n_samples, name='gen'):
with tf.variable_scope(name):
noise = tf.random_normal([n_samples, GEN_DIM])
output01 = tf_utils.linear(noise, 3 * DIM, name='fc-1')
output01 = tf_utils.relu(output01, name='relu-1')
output02 = tf_utils.linear(output01, 2 * DIM, name='fc-2')
output02 = tf_utils.relu(output02, name='relu-2')
output03 = tf_utils.linear(output02, DIM, name='fc-3')
output03 = tf_utils.relu(output03, name='relu-3')
output04 = tf_utils.linear(output03, 2 * DIM, name='fc-4')
output04 = tf_utils.relu(output04, name='relu-4')
output05 = tf_utils.linear(output04, 3 * DIM, name='fc-5')
output05 = tf_utils.relu(output05, name='relu-5')
output06 = tf_utils.linear(output05, GEN_DIM, name='fc-6')
# Reminder: a logit can be modeled as a linear function of the predictors
output05 = tf.nn.softmax(output06, name='softmax-1')
return output05
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 Discriminator(inputs, is_reuse=True, name='disc'):
with tf.variable_scope(name, reuse=is_reuse):
print('is_reuse: {}'.format(is_reuse))
output01 = tf_utils.linear(inputs, DIM, name='fc-1')
output01 = tf_utils.relu(output01, name='relu-1')
output02 = tf_utils.linear(output01, DIM, name='fc-2')
output02 = tf_utils.relu(output02, name='relu-2')
output03 = tf_utils.linear(output02, DIM, name='fc-3')
output03 = tf_utils.relu(output03, name='relu-3')
output04 = tf_utils.linear(output03, DIS_DIM, name='fc-4')
return output04
def network(self, inputs, name=None):
with tf.variable_scope(name):
tf_utils.print_activations(inputs)
# input of main reccurent layers
output = tf_utils.conv2d_mask(inputs,
2 * self.hidden_dims, [7, 7],
mask_type="A",
name='inputConv1')
# main recurrent layers
if self.flags.model == 'pixelcnn':
for idx in range(self.recurrent_length):
output = tf_utils.conv2d_mask(
output,
self.hidden_dims, [3, 3],
mask_type="B",
name='mainConv{}'.format(idx + 2))
output = tf_utils.relu(output,
name='mainRelu{}'.format(idx + 2))
elif self.flags.model == 'diagonal_bilstm':
for idx in range(self.recurrent_length):
output = self.diagonal_bilstm(output,
name='BiLSTM{}'.format(idx +
2))
elif self.flags.model == 'row_lstm':
raise NotImplementedError
else:
raise NotImplementedError
# output recurrent layers
for idx in range(self.out_recurrent_length):
output = tf_utils.conv2d_mask(output,
self.hidden_dims, [1, 1],
mask_type="B",
name='outputConv{}'.format(idx +
1))
output = tf_utils.relu(output,
name='outputRelu{}'.format(idx + 1))
# TODO: for color images, implement a 256-way softmax for each RGB channel here
output = tf_utils.conv2d_mask(output,
self.img_size[2], [1, 1],
mask_type="B",
name='outputConv3')
# output = tf_utils.sigmoid(output_logits, name='output_sigmoid')
return tf_utils.sigmoid(output), output
def Generator(n_samples, real_data_, name='gen'):
if FIXED_GENERATOR:
return real_data_ + (1. * tf.random_normal(tf.shape(real_data_)))
else:
with tf.variable_scope(name):
noise = tf.random_normal([n_samples, 100])
output01 = tf_utils.linear(noise, DIM, name='fc-1')
output01 = tf_utils.relu(output01, name='relu-1')
output02 = tf_utils.linear(output01, DIM, name='fc-2')
output02 = tf_utils.relu(output02, name='relu-2')
output03 = tf_utils.linear(output02, DIM, name='fc-3')
output03 = tf_utils.relu(output03, name='relu-3')
output04 = tf_utils.linear(output03, GEN_DIM, name='fc-4')
return output04
def Generator(n_samples, real_data_, name='gen'):
if FIXED_GENERATOR:
return real_data_ + (1. * tf.random_normal(tf.shape(real_data_)))
else:
with tf.variable_scope(name):
noise = tf.random_normal([n_samples, GEN_DIM])
output01 = tf_utils.linear(noise, DIM, name='fc-1')
output01 = tf_utils.relu(output01, name='relu-1')
output02 = tf_utils.linear(output01, DIM, name='fc-2')
output02 = tf_utils.relu(output02, name='relu-2')
output03 = tf_utils.linear(output02, DIM, name='fc-3')
output03 = tf_utils.relu(output03, name='relu-3')
output04 = tf_utils.linear(output03, GEN_DIM, name='fc-4')
# Reminder: a logit can be modeled as a linear function of the predictors
output05 = tf.nn.softmax(output04, name='softmax-1')
return output05
def forward_network(self, input_img, padding='SAME', reuse=False):
with tf.compat.v1.variable_scope(self.name, reuse=reuse):
tf_utils.print_activations(input_img, logger=self.logger)
inputs = self.conv2d_fixed_padding(inputs=input_img,
filters=64,
kernel_size=7,
strides=2,
name='conv1')
inputs = tf_utils.max_pool(inputs,
name='3x3_maxpool',
ksize=[1, 3, 3, 1],
strides=[1, 2, 2, 1],
logger=self.logger)
inputs = self.block_layer(inputs=inputs,
filters=64,
block_fn=self.bottleneck_block,
blocks=self.layers[0],
strides=1,
train_mode=self.train_mode,
name='block_layer1')
inputs = self.block_layer(inputs=inputs,
filters=128,
block_fn=self.bottleneck_block,
blocks=self.layers[1],
strides=2,
train_mode=self.train_mode,
name='block_layer2')
inputs = self.block_layer(inputs=inputs,
filters=256,
block_fn=self.bottleneck_block,
blocks=self.layers[2],
strides=2,
train_mode=self.train_mode,
name='block_layer3')
inputs = self.block_layer(inputs=inputs,
filters=512,
block_fn=self.bottleneck_block,
blocks=self.layers[3],
strides=2,
train_mode=self.train_mode,
name='block_layer4')
inputs = tf_utils.norm(inputs,
name='before_gap_batch_norm',
_type='batch',
_ops=self._ops,
is_train=self.train_mode,
logger=self.logger)
inputs = tf_utils.relu(inputs,
name='before_gap_relu',
logger=self.logger)
_, h, w, _ = inputs.get_shape().as_list()
inputs = tf_utils.avg_pool(inputs,
name='gap',
ksize=[1, h, w, 1],
strides=[1, 1, 1, 1],
logger=self.logger)
inputs = tf_utils.flatten(inputs,
name='flatten',
logger=self.logger)
logits = tf_utils.linear(inputs, self.num_classes, name='logits')
return logits
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
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 forward_network(self, img, padding='SAME', reuse=False):
with tf.compat.v1.variable_scope(self.name, reuse=reuse):
# Stage 0
s0_conv1 = tf_utils.conv2d(x=img,
output_dim=self.conv_dims[0],
k_h=3,
k_w=3,
d_h=1,
d_w=1,
padding=padding,
initializer='He',
name='s0_conv1')
s0_conv1 = tf_utils.relu(s0_conv1, name='relu_s0_conv1')
s0_conv2 = tf_utils.conv2d(x=s0_conv1,
output_dim=self.conv_dims[0],
k_h=3,
k_w=3,
d_h=1,
d_w=1,
padding=padding,
initializer='He',
name='s0_conv2')
s0_conv2 = tf_utils.norm(s0_conv2,
name='s0_norm1',
_type='batch',
_ops=self._ops,
is_train=False)
s0_conv2 = tf_utils.relu(s0_conv2, name='relu_s0_conv2')
# Stage 1
s1_maxpool = tf_utils.max_pool(x=s0_conv2, name='s1_maxpool2d')
s1_conv1 = tf_utils.conv2d(x=s1_maxpool,
output_dim=self.conv_dims[0],
k_h=3,
k_w=3,
d_h=1,
d_w=1,
padding=padding,
initializer='He',
name='s1_conv1')
s1_conv1 = tf_utils.norm(s1_conv1,
name='s1_norm0',
_type='batch',
_ops=self._ops,
is_train=False)
s1_conv1 = tf_utils.relu(s1_conv1, name='relu_s1_conv1')
s1_conv2 = tf_utils.conv2d(x=s1_conv1,
output_dim=self.conv_dims[1],
k_h=3,
k_w=3,
d_h=1,
d_w=1,
padding=padding,
initializer='He',
name='s1_conv2')
s1_conv2 = tf_utils.norm(s1_conv2,
name='s1_norm1',
_type='batch',
_ops=self._ops,
is_train=False)
s1_conv2 = tf_utils.relu(s1_conv2, name='relu_s1_conv2')
# Stage 2
s2_maxpool = tf_utils.max_pool(x=s1_conv2, name='s2_maxpool2d')
s2_conv1 = tf_utils.conv2d(x=s2_maxpool,
output_dim=self.conv_dims[2],
k_h=3,
k_w=3,
d_h=1,
d_w=1,
padding=padding,
initializer='He',
name='s2_conv1')
s2_conv1 = tf_utils.norm(s2_conv1,
name='s2_norm0',
_type='batch',
_ops=self._ops,
is_train=False)
s2_conv1 = tf_utils.relu(s2_conv1, name='relu_s2_conv1')
s2_conv2 = tf_utils.conv2d(x=s2_conv1,
output_dim=self.conv_dims[3],
k_h=3,
k_w=3,
d_h=1,
d_w=1,
padding=padding,
initializer='He',
name='s2_conv2')
s2_conv2 = tf_utils.norm(s2_conv2,
name='s2_norm1',
_type='batch',
_ops=self._ops,
is_train=False)
s2_conv2 = tf_utils.relu(s2_conv2, name='relu_s2_conv2')
# Stage 3
s3_maxpool = tf_utils.max_pool(x=s2_conv2, name='s3_maxpool2d')
s3_conv1 = tf_utils.conv2d(x=s3_maxpool,
output_dim=self.conv_dims[4],
k_h=3,
k_w=3,
d_h=1,
d_w=1,
padding=padding,
initializer='He',
name='s3_conv1')
s3_conv1 = tf_utils.norm(s3_conv1,
name='s3_norm0',
_type='batch',
_ops=self._ops,
is_train=False)
s3_conv1 = tf_utils.relu(s3_conv1, name='relu_s3_conv1')
s3_conv2 = tf_utils.conv2d(x=s3_conv1,
output_dim=self.conv_dims[5],
k_h=3,
k_w=3,
d_h=1,
d_w=1,
padding=padding,
initializer='He',
name='s3_conv2')
s3_conv2 = tf_utils.norm(s3_conv2,
name='s3_norm1',
_type='batch',
_ops=self._ops,
is_train=False)
s3_conv2 = tf_utils.relu(s3_conv2, name='relu_s3_conv2')
# Stage 4
s4_maxpool = tf_utils.max_pool(x=s3_conv2, name='s4_maxpool2d')
s4_conv1 = tf_utils.conv2d(x=s4_maxpool,
output_dim=self.conv_dims[6],
k_h=3,
k_w=3,
d_h=1,
d_w=1,
padding=padding,
initializer='He',
name='s4_conv1')
s4_conv1 = tf_utils.norm(s4_conv1,
name='s4_norm0',
_type='batch',
_ops=self._ops,
is_train=False)
s4_conv1 = tf_utils.relu(s4_conv1, name='relu_s4_conv1')
s4_conv2 = tf_utils.conv2d(x=s4_conv1,
output_dim=self.conv_dims[7],
k_h=3,
k_w=3,
d_h=1,
d_w=1,
padding=padding,
initializer='He',
name='s4_conv2')
s4_conv2 = tf_utils.norm(s4_conv2,
name='s4_norm1',
_type='batch',
_ops=self._ops,
is_train=False)
s4_conv2 = tf_utils.relu(s4_conv2, name='relu_s4_conv2')
s4_conv2_drop = tf_utils.dropout(x=s4_conv2,
keep_prob=0.,
name='s4_dropout')
# Stage 5
s5_maxpool = tf_utils.max_pool(x=s4_conv2_drop,
name='s5_maxpool2d')
s5_conv1 = tf_utils.conv2d(x=s5_maxpool,
output_dim=self.conv_dims[8],
k_h=3,
k_w=3,
d_h=1,
d_w=1,
padding=padding,
initializer='He',
name='s5_conv1')
s5_conv1 = tf_utils.norm(s5_conv1,
name='s5_norm0',
_type='batch',
_ops=self._ops,
is_train=False)
s5_conv1 = tf_utils.relu(s5_conv1, name='relu_s5_conv1')
s5_conv2 = tf_utils.conv2d(x=s5_conv1,
output_dim=self.conv_dims[9],
k_h=3,
k_w=3,
d_h=1,
d_w=1,
padding=padding,
initializer='He',
name='s5_conv2')
s5_conv2 = tf_utils.norm(s5_conv2,
name='s5_norm1',
_type='batch',
_ops=self._ops,
is_train=False)
s5_conv2 = tf_utils.relu(s5_conv2, name='relu_s5_conv2')
s5_conv2_drop = tf_utils.dropout(x=s5_conv2,
keep_prob=0.,
name='s5_dropout')
# Stage 6
s6_deconv1 = tf_utils.deconv2d(x=s5_conv2_drop,
output_dim=self.conv_dims[10],
k_h=2,
k_w=2,
initializer='He',
name='s6_deconv1')
s6_deconv1 = tf_utils.norm(s6_deconv1,
name='s6_norm0',
_type='batch',
_ops=self._ops,
is_train=False)
s6_deconv1 = tf_utils.relu(s6_deconv1, name='relu_s6_deconv1')
# Cropping
w1 = s4_conv2_drop.get_shape().as_list()[2]
w2 = s6_deconv1.get_shape().as_list()[2] - s4_conv2_drop.get_shape(
).as_list()[2]
s6_deconv1_split, _ = tf.split(s6_deconv1,
num_or_size_splits=[w1, w2],
axis=2,
name='axis2_split')
tf_utils.print_activations(s6_deconv1_split)
# Concat
s6_concat = tf_utils.concat(
values=[s6_deconv1_split, s4_conv2_drop],
axis=3,
name='s6_axis3_concat')
s6_conv2 = tf_utils.conv2d(x=s6_concat,
output_dim=self.conv_dims[11],
k_h=3,
k_w=3,
d_h=1,
d_w=1,
padding=padding,
initializer='He',
name='s6_conv2')
s6_conv2 = tf_utils.norm(s6_conv2,
name='s6_norm1',
_type='batch',
_ops=self._ops,
is_train=False)
s6_conv2 = tf_utils.relu(s6_conv2, name='relu_s6_conv2')
s6_conv3 = tf_utils.conv2d(x=s6_conv2,
output_dim=self.conv_dims[12],
k_h=3,
k_w=3,
d_h=1,
d_w=1,
padding=padding,
initializer='He',
name='s6_conv3')
s6_conv3 = tf_utils.norm(s6_conv3,
name='s6_norm2',
_type='batch',
_ops=self._ops,
is_train=False)
s6_conv3 = tf_utils.relu(s6_conv3, name='relu_s6_conv3')
# Stage 7
s7_deconv1 = tf_utils.deconv2d(x=s6_conv3,
output_dim=self.conv_dims[13],
k_h=2,
k_w=2,
initializer='He',
name='s7_deconv1')
s7_deconv1 = tf_utils.norm(s7_deconv1,
name='s7_norm0',
_type='batch',
_ops=self._ops,
is_train=False)
s7_deconv1 = tf_utils.relu(s7_deconv1, name='relu_s7_deconv1')
# Concat
s7_concat = tf_utils.concat(values=[s7_deconv1, s3_conv2],
axis=3,
name='s7_axis3_concat')
s7_conv2 = tf_utils.conv2d(x=s7_concat,
output_dim=self.conv_dims[14],
k_h=3,
k_w=3,
d_h=1,
d_w=1,
padding=padding,
initializer='He',
name='s7_conv2')
s7_conv2 = tf_utils.norm(s7_conv2,
name='s7_norm1',
_type='batch',
_ops=self._ops,
is_train=False)
s7_conv2 = tf_utils.relu(s7_conv2, name='relu_s7_conv2')
s7_conv3 = tf_utils.conv2d(x=s7_conv2,
output_dim=self.conv_dims[15],
k_h=3,
k_w=3,
d_h=1,
d_w=1,
padding=padding,
initializer='He',
name='s7_conv3')
s7_conv3 = tf_utils.norm(s7_conv3,
name='s7_norm2',
_type='batch',
_ops=self._ops,
is_train=False)
s7_conv3 = tf_utils.relu(s7_conv3, name='relu_s7_conv3')
# Stage 8
s8_deconv1 = tf_utils.deconv2d(x=s7_conv3,
output_dim=self.conv_dims[16],
k_h=2,
k_w=2,
initializer='He',
name='s8_deconv1')
s8_deconv1 = tf_utils.norm(s8_deconv1,
name='s8_norm0',
_type='batch',
_ops=self._ops,
is_train=False)
s8_deconv1 = tf_utils.relu(s8_deconv1, name='relu_s8_deconv1')
# Concat
s8_concat = tf_utils.concat(values=[s8_deconv1, s2_conv2],
axis=3,
name='s8_axis3_concat')
s8_conv2 = tf_utils.conv2d(x=s8_concat,
output_dim=self.conv_dims[17],
k_h=3,
k_w=3,
d_h=1,
d_w=1,
padding=padding,
initializer='He',
name='s8_conv2')
s8_conv2 = tf_utils.norm(s8_conv2,
name='s8_norm1',
_type='batch',
_ops=self._ops,
is_train=False)
s8_conv2 = tf_utils.relu(s8_conv2, name='relu_s8_conv2')
s8_conv3 = tf_utils.conv2d(x=s8_conv2,
output_dim=self.conv_dims[18],
k_h=3,
k_w=3,
d_h=1,
d_w=1,
padding=padding,
initializer='He',
name='s8_conv3')
s8_conv3 = tf_utils.norm(s8_conv3,
name='s8_norm2',
_type='batch',
_ops=self._ops,
is_train=False)
s8_conv3 = tf_utils.relu(s8_conv3, name='relu_conv3')
# Stage 9
s9_deconv1 = tf_utils.deconv2d(x=s8_conv3,
output_dim=self.conv_dims[19],
k_h=2,
k_w=2,
initializer='He',
name='s9_deconv1')
s9_deconv1 = tf_utils.norm(s9_deconv1,
name='s9_norm0',
_type='batch',
_ops=self._ops,
is_train=False)
s9_deconv1 = tf_utils.relu(s9_deconv1, name='relu_s9_deconv1')
# Concat
s9_concat = tf_utils.concat(values=[s9_deconv1, s1_conv2],
axis=3,
name='s9_axis3_concat')
s9_conv2 = tf_utils.conv2d(x=s9_concat,
output_dim=self.conv_dims[20],
k_h=3,
k_w=3,
d_h=1,
d_w=1,
padding=padding,
initializer='He',
name='s9_conv2')
s9_conv2 = tf_utils.norm(s9_conv2,
name='s9_norm1',
_type='batch',
_ops=self._ops,
is_train=False)
s9_conv2 = tf_utils.relu(s9_conv2, name='relu_s9_conv2')
s9_conv3 = tf_utils.conv2d(x=s9_conv2,
output_dim=self.conv_dims[21],
k_h=3,
k_w=3,
d_h=1,
d_w=1,
padding=padding,
initializer='He',
name='s9_conv3')
s9_conv3 = tf_utils.norm(s9_conv3,
name='s9_norm2',
_type='batch',
_ops=self._ops,
is_train=False)
s9_conv3 = tf_utils.relu(s9_conv3, name='relu_s9_conv3')
s10_deconv1 = tf_utils.deconv2d(x=s9_conv3,
output_dim=self.conv_dims[-1],
k_h=2,
k_w=2,
initializer='He',
name='s10_deconv1')
s10_deconv1 = tf_utils.norm(s10_deconv1,
name='s10_norm0',
_type='batch',
_ops=self._ops,
is_train=False)
s10_deconv1 = tf_utils.relu(s10_deconv1, name='relu_s10_deconv1')
# Concat
s10_concat = tf_utils.concat(values=[s10_deconv1, s0_conv2],
axis=3,
name='s10_axis3_concat')
s10_conv2 = tf_utils.conv2d(s10_concat,
output_dim=self.conv_dims[-1],
k_h=3,
k_w=3,
d_h=1,
d_w=1,
padding=padding,
initializer='He',
name='s10_conv2')
s10_conv2 = tf_utils.norm(s10_conv2,
name='s10_norm1',
_type='batch',
_ops=self._ops,
is_train=False)
s10_conv2 = tf_utils.relu(s10_conv2, name='relu_s10_conv2')
s10_conv3 = tf_utils.conv2d(x=s10_conv2,
output_dim=self.conv_dims[-1],
k_h=3,
k_w=3,
d_h=1,
d_w=1,
padding=padding,
initializer='He',
name='s10_conv3')
s10_conv3 = tf_utils.norm(s10_conv3,
name='s10_norm2',
_type='batch',
_ops=self._ops,
is_train=False)
s10_conv3 = tf_utils.relu(s10_conv3, name='relu_s10_conv3')
output = tf_utils.conv2d(s10_conv3,
output_dim=self.num_classes,
k_h=1,
k_w=1,
d_h=1,
d_w=1,
padding=padding,
initializer='He',
name='output')
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):
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 __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 forward_network(self, inputImg, padding='SAME', reuse=False):
with tf.compat.v1.variable_scope(self.name, reuse=reuse):
# This part is for compatible between input size [640, 400] and [320, 200]
if self.resize_factor == 1.0:
# Stage 0
tf_utils.print_activations(inputImg, logger=self.logger)
s0_conv1 = tf_utils.conv2d(x=inputImg, output_dim=self.conv_dims[0], k_h=3, k_w=3, d_h=1, d_w=1,
padding=padding, initializer='He', name='s0_conv1', logger=self.logger)
s0_conv1 = tf_utils.relu(s0_conv1, name='relu_s0_conv1', logger=self.logger)
s0_conv2 = tf_utils.conv2d(x=s0_conv1, output_dim=self.conv_dims[0], k_h=3, k_w=3, d_h=1, d_w=1,
padding=padding, initializer='He', name='s0_conv2', logger=self.logger)
if self.use_batch_norm:
s0_conv2 = tf_utils.norm(s0_conv2, name='s0_norm1', _type='batch', _ops=self._ops,
is_train=self.trainMode, logger=self.logger)
s0_conv2 = tf_utils.relu(s0_conv2, name='relu_s0_conv2', logger=self.logger)
# Stage 1
s1_maxpool = tf_utils.max_pool(x=s0_conv2, name='s1_maxpool2d', logger=self.logger)
s1_conv1 = tf_utils.conv2d(x=s1_maxpool, output_dim=self.conv_dims[0], k_h=3, k_w=3, d_h=1, d_w=1,
padding=padding, initializer='He', name='s1_conv1', logger=self.logger)
if self.use_batch_norm:
s1_conv1 = tf_utils.norm(s1_conv1, name='s1_norm0', _type='batch', _ops=self._ops,
is_train=self.trainMode, logger=self.logger)
s1_conv1 = tf_utils.relu(s1_conv1, name='relu_s1_conv1', logger=self.logger)
s1_conv2 = tf_utils.conv2d(x=s1_conv1, output_dim=self.conv_dims[1], k_h=3, k_w=3, d_h=1, d_w=1,
padding=padding, initializer='He', name='s1_conv2', logger=self.logger)
if self.use_batch_norm:
s1_conv2 = tf_utils.norm(s1_conv2, name='s1_norm1', _type='batch', _ops=self._ops,
is_train=self.trainMode, logger=self.logger)
s1_conv2 = tf_utils.relu(s1_conv2, name='relu_s1_conv2', logger=self.logger)
else:
# Stage 1
tf_utils.print_activations(inputImg, logger=self.logger)
s1_conv1 = tf_utils.conv2d(x=inputImg, output_dim=self.conv_dims[0], k_h=3, k_w=3, d_h=1, d_w=1,
padding=padding, initializer='He', name='s1_conv1', logger=self.logger)
s1_conv1 = tf_utils.relu(s1_conv1, name='relu_s1_conv1', logger=self.logger)
s1_conv2 = tf_utils.conv2d(x=s1_conv1, output_dim=self.conv_dims[1], k_h=3, k_w=3, d_h=1, d_w=1,
padding=padding, initializer='He', name='s1_conv2', logger=self.logger)
if self.use_batch_norm:
s1_conv2 = tf_utils.norm(s1_conv2, name='s1_norm1', _type='batch', _ops=self._ops,
is_train=self.trainMode, logger=self.logger)
s1_conv2 = tf_utils.relu(s1_conv2, name='relu_s1_conv2', logger=self.logger)
# Stage 2
s2_maxpool = tf_utils.max_pool(x=s1_conv2, name='s2_maxpool2d', logger=self.logger)
s2_conv1 = tf_utils.conv2d(x=s2_maxpool, output_dim=self.conv_dims[2], k_h=3, k_w=3, d_h=1, d_w=1,
padding=padding, initializer='He', name='s2_conv1', logger=self.logger)
if self.use_batch_norm:
s2_conv1 = tf_utils.norm(s2_conv1, name='s2_norm0', _type='batch', _ops=self._ops,
is_train=self.trainMode, logger=self.logger)
s2_conv1 = tf_utils.relu(s2_conv1, name='relu_s2_conv1', logger=self.logger)
s2_conv2 = tf_utils.conv2d(x=s2_conv1, output_dim=self.conv_dims[3], k_h=3, k_w=3, d_h=1, d_w=1,
padding=padding, initializer='He', name='s2_conv2', logger=self.logger)
if self.use_batch_norm:
s2_conv2 = tf_utils.norm(s2_conv2, name='s2_norm1', _type='batch', _ops=self._ops,
is_train=self.trainMode, logger=self.logger)
s2_conv2 = tf_utils.relu(s2_conv2, name='relu_s2_conv2', logger=self.logger)
# Stage 3
s3_maxpool = tf_utils.max_pool(x=s2_conv2, name='s3_maxpool2d', logger=self.logger)
s3_conv1 = tf_utils.conv2d(x=s3_maxpool, output_dim=self.conv_dims[4], k_h=3, k_w=3, d_h=1, d_w=1,
padding=padding, initializer='He', name='s3_conv1', logger=self.logger)
if self.use_batch_norm:
s3_conv1 = tf_utils.norm(s3_conv1, name='s3_norm0', _type='batch', _ops=self._ops,
is_train=self.trainMode, logger=self.logger)
s3_conv1 = tf_utils.relu(s3_conv1, name='relu_s3_conv1', logger=self.logger)
s3_conv2 = tf_utils.conv2d(x=s3_conv1, output_dim=self.conv_dims[5], k_h=3, k_w=3, d_h=1, d_w=1,
padding=padding, initializer='He', name='s3_conv2', logger=self.logger)
if self.use_batch_norm:
s3_conv2 = tf_utils.norm(s3_conv2, name='s3_norm1', _type='batch', _ops=self._ops,
is_train=self.trainMode, logger=self.logger)
s3_conv2 = tf_utils.relu(s3_conv2, name='relu_s3_conv2', logger=self.logger)
# Stage 4
s4_maxpool = tf_utils.max_pool(x=s3_conv2, name='s4_maxpool2d', logger=self.logger)
s4_conv1 = tf_utils.conv2d(x=s4_maxpool, output_dim=self.conv_dims[6], k_h=3, k_w=3, d_h=1, d_w=1,
padding=padding, initializer='He', name='s4_conv1', logger=self.logger)
if self.use_batch_norm:
s4_conv1 = tf_utils.norm(s4_conv1, name='s4_norm0', _type='batch', _ops=self._ops,
is_train=self.trainMode, logger=self.logger)
s4_conv1 = tf_utils.relu(s4_conv1, name='relu_s4_conv1', logger=self.logger)
s4_conv2 = tf_utils.conv2d(x=s4_conv1, output_dim=self.conv_dims[7], k_h=3, k_w=3, d_h=1, d_w=1,
padding=padding, initializer='He', name='s4_conv2', logger=self.logger)
if self.use_batch_norm:
s4_conv2 = tf_utils.norm(s4_conv2, name='s4_norm1', _type='batch', _ops=self._ops,
is_train=self.trainMode, logger=self.logger)
s4_conv2 = tf_utils.relu(s4_conv2, name='relu_s4_conv2', logger=self.logger)
s4_conv2_drop = tf_utils.dropout(x=s4_conv2, keep_prob=self.ratePh, name='s4_dropout',
logger=self.logger)
# Stage 5
s5_maxpool = tf_utils.max_pool(x=s4_conv2_drop, name='s5_maxpool2d', logger=self.logger)
s5_conv1 = tf_utils.conv2d(x=s5_maxpool, output_dim=self.conv_dims[8], k_h=3, k_w=3, d_h=1, d_w=1,
padding=padding, initializer='He', name='s5_conv1', logger=self.logger)
if self.use_batch_norm:
s5_conv1 = tf_utils.norm(s5_conv1, name='s5_norm0', _type='batch', _ops=self._ops,
is_train=self.trainMode, logger=self.logger)
s5_conv1 = tf_utils.relu(s5_conv1, name='relu_s5_conv1', logger=self.logger)
s5_conv2 = tf_utils.conv2d(x=s5_conv1, output_dim=self.conv_dims[9], k_h=3, k_w=3, d_h=1, d_w=1,
padding=padding, initializer='He', name='s5_conv2', logger=self.logger)
if self.use_batch_norm:
s5_conv2 = tf_utils.norm(s5_conv2, name='s5_norm1', _type='batch', _ops=self._ops,
is_train=self.trainMode, logger=self.logger)
s5_conv2 = tf_utils.relu(s5_conv2, name='relu_s5_conv2', logger=self.logger)
s5_conv2_drop = tf_utils.dropout(x=s5_conv2, keep_prob=self.ratePh, name='s5_dropout',
logger=self.logger)
# Stage 6
s6_deconv1 = tf_utils.deconv2d(x=s5_conv2_drop, output_dim=self.conv_dims[10], k_h=2, k_w=2,
initializer='He', name='s6_deconv1', logger=self.logger)
if self.use_batch_norm:
s6_deconv1 = tf_utils.norm(s6_deconv1, name='s6_norm0', _type='batch', _ops=self._ops,
is_train=self.trainMode, logger=self.logger)
s6_deconv1 = tf_utils.relu(s6_deconv1, name='relu_s6_deconv1', logger=self.logger)
# Cropping
w1 = s4_conv2_drop.get_shape().as_list()[2]
w2 = s6_deconv1.get_shape().as_list()[2] - s4_conv2_drop.get_shape().as_list()[2]
s6_deconv1_split, _ = tf.split(s6_deconv1, num_or_size_splits=[w1, w2], axis=2, name='axis2_split')
tf_utils.print_activations(s6_deconv1_split, logger=self.logger)
# Concat
s6_concat = tf_utils.concat(values=[s6_deconv1_split, s4_conv2_drop], axis=3, name='s6_axis3_concat',
logger=self.logger)
s6_conv2 = tf_utils.conv2d(x=s6_concat, output_dim=self.conv_dims[11], k_h=3, k_w=3, d_h=1, d_w=1,
padding=padding, initializer='He', name='s6_conv2', logger=self.logger)
if self.use_batch_norm:
s6_conv2 = tf_utils.norm(s6_conv2, name='s6_norm1', _type='batch', _ops=self._ops,
is_train=self.trainMode, logger=self.logger)
s6_conv2 = tf_utils.relu(s6_conv2, name='relu_s6_conv2', logger=self.logger)
s6_conv3 = tf_utils.conv2d(x=s6_conv2, output_dim=self.conv_dims[12], k_h=3, k_w=3, d_h=1, d_w=1,
padding=padding, initializer='He', name='s6_conv3', logger=self.logger)
if self.use_batch_norm:
s6_conv3 = tf_utils.norm(s6_conv3, name='s6_norm2', _type='batch', _ops=self._ops,
is_train=self.trainMode, logger=self.logger)
s6_conv3 = tf_utils.relu(s6_conv3, name='relu_s6_conv3', logger=self.logger)
# Stage 7
s7_deconv1 = tf_utils.deconv2d(x=s6_conv3, output_dim=self.conv_dims[13], k_h=2, k_w=2, initializer='He',
name='s7_deconv1', logger=self.logger)
if self.use_batch_norm:
s7_deconv1 = tf_utils.norm(s7_deconv1, name='s7_norm0', _type='batch', _ops=self._ops,
is_train=self.trainMode, logger=self.logger)
s7_deconv1 = tf_utils.relu(s7_deconv1, name='relu_s7_deconv1', logger=self.logger)
# Concat
s7_concat = tf_utils.concat(values=[s7_deconv1, s3_conv2], axis=3, name='s7_axis3_concat',
logger=self.logger)
s7_conv2 = tf_utils.conv2d(x=s7_concat, output_dim=self.conv_dims[14], k_h=3, k_w=3, d_h=1, d_w=1,
padding=padding, initializer='He', name='s7_conv2', logger=self.logger)
if self.use_batch_norm:
s7_conv2 = tf_utils.norm(s7_conv2, name='s7_norm1', _type='batch', _ops=self._ops,
is_train=self.trainMode, logger=self.logger)
s7_conv2 = tf_utils.relu(s7_conv2, name='relu_s7_conv2', logger=self.logger)
s7_conv3 = tf_utils.conv2d(x=s7_conv2, output_dim=self.conv_dims[15], k_h=3, k_w=3, d_h=1, d_w=1,
padding=padding, initializer='He', name='s7_conv3', logger=self.logger)
if self.use_batch_norm:
s7_conv3 = tf_utils.norm(s7_conv3, name='s7_norm2', _type='batch', _ops=self._ops,
is_train=self.trainMode, logger=self.logger)
s7_conv3 = tf_utils.relu(s7_conv3, name='relu_s7_conv3', logger=self.logger)
# Stage 8
s8_deconv1 = tf_utils.deconv2d(x=s7_conv3, output_dim=self.conv_dims[16], k_h=2, k_w=2, initializer='He',
name='s8_deconv1', logger=self.logger)
if self.use_batch_norm:
s8_deconv1 = tf_utils.norm(s8_deconv1, name='s8_norm0', _type='batch', _ops=self._ops,
is_train=self.trainMode, logger=self.logger)
s8_deconv1 = tf_utils.relu(s8_deconv1, name='relu_s8_deconv1', logger=self.logger)
# Concat
s8_concat = tf_utils.concat(values=[s8_deconv1,s2_conv2], axis=3, name='s8_axis3_concat',
logger=self.logger)
s8_conv2 = tf_utils.conv2d(x=s8_concat, output_dim=self.conv_dims[17], k_h=3, k_w=3, d_h=1, d_w=1,
padding=padding, initializer='He', name='s8_conv2', logger=self.logger)
if self.use_batch_norm:
s8_conv2 = tf_utils.norm(s8_conv2, name='s8_norm1', _type='batch', _ops=self._ops,
is_train=self.trainMode, logger=self.logger)
s8_conv2 = tf_utils.relu(s8_conv2, name='relu_s8_conv2', logger=self.logger)
s8_conv3 = tf_utils.conv2d(x=s8_conv2, output_dim=self.conv_dims[18], k_h=3, k_w=3, d_h=1, d_w=1,
padding=padding, initializer='He', name='s8_conv3', logger=self.logger)
if self.use_batch_norm:
s8_conv3 = tf_utils.norm(s8_conv3, name='s8_norm2', _type='batch', _ops=self._ops,
is_train=self.trainMode, logger=self.logger)
s8_conv3 = tf_utils.relu(s8_conv3, name='relu_conv3', logger=self.logger)
# Stage 9
s9_deconv1 = tf_utils.deconv2d(x=s8_conv3, output_dim=self.conv_dims[19], k_h=2, k_w=2,
initializer='He', name='s9_deconv1', logger=self.logger)
if self.use_batch_norm:
s9_deconv1 = tf_utils.norm(s9_deconv1, name='s9_norm0', _type='batch', _ops=self._ops,
is_train=self.trainMode, logger=self.logger)
s9_deconv1 = tf_utils.relu(s9_deconv1, name='relu_s9_deconv1', logger=self.logger)
# Concat
s9_concat = tf_utils.concat(values=[s9_deconv1, s1_conv2], axis=3, name='s9_axis3_concat',
logger=self.logger)
s9_conv2 = tf_utils.conv2d(x=s9_concat, output_dim=self.conv_dims[20], k_h=3, k_w=3, d_h=1, d_w=1,
padding=padding, initializer='He', name='s9_conv2', logger=self.logger)
if self.use_batch_norm:
s9_conv2 = tf_utils.norm(s9_conv2, name='s9_norm1', _type='batch', _ops=self._ops,
is_train=self.trainMode, logger=self.logger)
s9_conv2 = tf_utils.relu(s9_conv2, name='relu_s9_conv2', logger=self.logger)
s9_conv3 = tf_utils.conv2d(x=s9_conv2, output_dim=self.conv_dims[21], k_h=3, k_w=3, d_h=1, d_w=1,
padding=padding, initializer='He', name='s9_conv3', logger=self.logger)
if self.use_batch_norm:
s9_conv3 = tf_utils.norm(s9_conv3, name='s9_norm2', _type='batch', _ops=self._ops,
is_train=self.trainMode, logger=self.logger)
s9_conv3 = tf_utils.relu(s9_conv3, name='relu_s9_conv3', logger=self.logger)
if self.resize_factor == 1.0:
s10_deconv1 = tf_utils.deconv2d(x=s9_conv3, output_dim=self.conv_dims[-1], k_h=2, k_w=2,
initializer='He', name='s10_deconv1', logger=self.logger)
if self.use_batch_norm:
s10_deconv1 = tf_utils.norm(s10_deconv1, name='s10_norm0', _type='batch', _ops=self._ops,
is_train=self.trainMode, logger=self.logger)
s10_deconv1 = tf_utils.relu(s10_deconv1, name='relu_s10_deconv1', logger=self.logger)
# Concat
s10_concat = tf_utils.concat(values=[s10_deconv1, s0_conv2], axis=3, name='s10_axis3_concat',
logger=self.logger)
s10_conv2 = tf_utils.conv2d(s10_concat, output_dim=self.conv_dims[-1], k_h=3, k_w=3, d_h=1, d_w=1,
padding=padding, initializer='He', name='s10_conv2', logger=self.logger)
if self.use_batch_norm:
s10_conv2 = tf_utils.norm(s10_conv2, name='s10_norm1', _type='batch', _ops=self._ops,
is_train=self.trainMode, logger=self.logger)
s10_conv2 = tf_utils.relu(s10_conv2, name='relu_s10_conv2', logger=self.logger)
s10_conv3 = tf_utils.conv2d(x=s10_conv2, output_dim=self.conv_dims[-1], k_h=3, k_w=3, d_h=1, d_w=1,
padding=padding, initializer='He', name='s10_conv3', logger=self.logger)
if self.use_batch_norm:
s10_conv3 = tf_utils.norm(s10_conv3, name='s10_norm2', _type='batch', _ops=self._ops,
is_train=self.trainMode, logger=self.logger)
s10_conv3 = tf_utils.relu(s10_conv3, name='relu_s10_conv3', logger=self.logger)
output = tf_utils.conv2d(s10_conv3, output_dim=self.numClasses, k_h=1, k_w=1, d_h=1, d_w=1,
padding=padding, initializer='He', name='output', logger=self.logger)
else:
output = tf_utils.conv2d(s9_conv3, output_dim=self.numClasses, k_h=1, k_w=1, d_h=1, d_w=1,
padding=padding, initializer='He', name='output', logger=self.logger)
return output
def __init__(self, input_dim, output_dim=[1000, 1000, 10], optimizer=None, use_dropout=True, lr=0.001,
weight_decay=1e-4, random_seed=123, is_train=True, log_dir=None, name=None):
self.name = name
self.is_train = is_train
self.log_dir = log_dir
self.cur_lr = None
self.logger, self.file_handler, self.stream_handler = utils.init_logger(log_dir=self.log_dir,
name=self.name,
is_train=self.is_train)
with tf.variable_scope(self.name):
# Placeholders for inputs
self.X = tf.placeholder(dtype=tf.float32, shape=[None, input_dim], name='X')
tf_utils.print_activations(self.X, logger=self.logger if self.is_train else None)
self.y = tf.placeholder(dtype=tf.float32, shape=[None, output_dim[-1]], name='y')
self.y_cls = tf.math.argmax(input=self.y, axis=1)
self.keep_prob = tf.placeholder(tf.float32, name='keep_prob')
# Placeholders for TensorBoard
self.train_acc = tf.placeholder(tf.float32, name='train_acc')
self.val_acc = tf.placeholder(tf.float32, name='val_acc')
net = self.X
for idx in range(len(output_dim) - 1):
net = tf_utils.linear(x=net,
output_size=output_dim[idx],
name='fc'+str(idx),
logger=self.logger if self.is_train else None)
if use_dropout:
net = tf_utils.dropout(x=net,
keep_prob=self.keep_prob,
seed=random_seed,
name='dropout'+str(idx),
logger=self.logger if self.is_train else None)
net = tf_utils.relu(x=net,
name='relu'+str(idx),
logger=self.logger if self.is_train else None)
# Last predict layer
self.y_pred = tf_utils.linear(net, output_size=output_dim[-1], name='last_fc')
tf_utils.print_activations(self.y_pred, logger=self.logger if self.is_train else None)
# Loss = data loss + regularization term
self.data_loss = tf.math.reduce_mean(
tf.nn.sigmoid_cross_entropy_with_logits(logits=self.y_pred, labels=self.y))
self.reg_term = weight_decay * tf.reduce_sum(
[tf.nn.l2_loss(weight) for weight in tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES)])
self.loss = self.data_loss + self.reg_term
# Optimizer
self.train_op, self.cur_lr = optimizer_fn(optimizer, lr=lr, loss=self.loss, name=self.name)
# Accuracy etc
self.y_pred_cls = tf.math.argmax(input=self.y_pred, axis=1)
correct_prediction = tf.math.equal(self.y_pred_cls, self.y_cls)
self.accuracy = tf.reduce_mean(tf.cast(correct_prediction, dtype=tf.float32)) * 100.
self._tensorboard()
tf_utils.show_all_variables(logger=self.logger if self.is_train else None)
def u_net(self):
# Stage 1
tf_utils.print_activations(self.inp_img, logger=self.logger)
s1_conv1 = tf_utils.conv2d(x=self.inp_img,
output_dim=self.conv_dims[0],
k_h=3,
k_w=3,
d_h=1,
d_w=1,
padding='VALID',
initializer='He',
name='s1_conv1',
logger=self.logger)
s1_conv1 = tf_utils.relu(s1_conv1,
name='relu_s1_conv1',
logger=self.logger)
s1_conv2 = tf_utils.conv2d(x=s1_conv1,
output_dim=self.conv_dims[1],
k_h=3,
k_w=3,
d_h=1,
d_w=1,
padding='VALID',
initializer='He',
name='s1_conv2',
logger=self.logger)
s1_conv2 = tf_utils.relu(s1_conv2,
name='relu_s1_conv2',
logger=self.logger)
# Stage 2
s2_maxpool = tf_utils.max_pool(x=s1_conv2,
name='s2_maxpool',
logger=self.logger)
s2_conv1 = tf_utils.conv2d(x=s2_maxpool,
output_dim=self.conv_dims[2],
k_h=3,
k_w=3,
d_h=1,
d_w=1,
padding='VALID',
initializer='He',
name='s2_conv1',
logger=self.logger)
s2_conv1 = tf_utils.relu(s2_conv1,
name='relu_s2_conv1',
logger=self.logger)
s2_conv2 = tf_utils.conv2d(x=s2_conv1,
output_dim=self.conv_dims[3],
k_h=3,
k_w=3,
d_h=1,
d_w=1,
padding='VALID',
initializer='He',
name='s2_conv2',
logger=self.logger)
s2_conv2 = tf_utils.relu(s2_conv2,
name='relu_s2_conv2',
logger=self.logger)
# Stage 3
s3_maxpool = tf_utils.max_pool(x=s2_conv2,
name='s3_maxpool',
logger=self.logger)
s3_conv1 = tf_utils.conv2d(x=s3_maxpool,
output_dim=self.conv_dims[4],
k_h=3,
k_w=3,
d_h=1,
d_w=1,
padding='VALID',
initializer='He',
name='s3_conv1',
logger=self.logger)
s3_conv1 = tf_utils.relu(s3_conv1,
name='relu_s3_conv1',
logger=self.logger)
s3_conv2 = tf_utils.conv2d(x=s3_conv1,
output_dim=self.conv_dims[5],
k_h=3,
k_w=3,
d_h=1,
d_w=1,
padding='VALID',
initializer='He',
name='s3_conv2',
logger=self.logger)
s3_conv2 = tf_utils.relu(s3_conv2,
name='relu_s3_conv2',
logger=self.logger)
# Stage 4
s4_maxpool = tf_utils.max_pool(x=s3_conv2,
name='s4_maxpool',
logger=self.logger)
s4_conv1 = tf_utils.conv2d(x=s4_maxpool,
output_dim=self.conv_dims[6],
k_h=3,
k_w=3,
d_h=1,
d_w=1,
padding='VALID',
initializer='He',
name='s4_conv1',
logger=self.logger)
s4_conv1 = tf_utils.relu(s4_conv1,
name='relu_s4_conv1',
logger=self.logger)
s4_conv2 = tf_utils.conv2d(x=s4_conv1,
output_dim=self.conv_dims[7],
k_h=3,
k_w=3,
d_h=1,
d_w=1,
padding='VALID',
initializer='He',
name='s4_conv2',
logger=self.logger)
s4_conv2 = tf_utils.relu(s4_conv2,
name='relu_s4_conv2',
logger=self.logger)
s4_conv2_drop = tf_utils.dropout(x=s4_conv2,
keep_prob=self.keep_prob,
name='s4_conv2_dropout',
logger=self.logger)
# Stage 5
s5_maxpool = tf_utils.max_pool(x=s4_conv2_drop,
name='s5_maxpool',
logger=self.logger)
s5_conv1 = tf_utils.conv2d(x=s5_maxpool,
output_dim=self.conv_dims[8],
k_h=3,
k_w=3,
d_h=1,
d_w=1,
padding='VALID',
initializer='He',
name='s5_conv1',
logger=self.logger)
s5_conv1 = tf_utils.relu(s5_conv1,
name='relu_s5_conv1',
logger=self.logger)
s5_conv2 = tf_utils.conv2d(x=s5_conv1,
output_dim=self.conv_dims[9],
k_h=3,
k_w=3,
d_h=1,
d_w=1,
padding='VALID',
initializer='He',
name='s5_conv2',
logger=self.logger)
s5_conv2 = tf_utils.relu(s5_conv2,
name='relu_s5_conv2',
logger=self.logger)
s5_conv2_drop = tf_utils.dropout(x=s5_conv2,
keep_prob=self.keep_prob,
name='s5_conv2_dropout',
logger=self.logger)
# Stage 6
s6_deconv1 = tf_utils.deconv2d(x=s5_conv2_drop,
output_dim=self.conv_dims[10],
k_h=2,
k_w=2,
initializer='He',
name='s6_deconv1',
logger=self.logger)
s6_deconv1 = tf_utils.relu(s6_deconv1,
name='relu_s6_deconv1',
logger=self.logger)
# Cropping
h1, w1 = s4_conv2_drop.get_shape().as_list()[1:3]
h2, w2 = s6_deconv1.get_shape().as_list()[1:3]
s4_conv2_crop = tf.image.crop_to_bounding_box(
image=s4_conv2_drop,
offset_height=int(0.5 * (h1 - h2)),
offset_width=int(0.5 * (w1 - w2)),
target_height=h2,
target_width=w2)
tf_utils.print_activations(s4_conv2_crop, logger=self.logger)
s6_concat = tf_utils.concat(values=[s4_conv2_crop, s6_deconv1],
axis=3,
name='s6_concat',
logger=self.logger)
s6_conv2 = tf_utils.conv2d(x=s6_concat,
output_dim=self.conv_dims[11],
k_h=3,
k_w=3,
d_h=1,
d_w=1,
padding='VALID',
initializer='He',
name='s6_conv2',
logger=self.logger)
s6_conv2 = tf_utils.relu(s6_conv2,
name='relu_s6_conv2',
logger=self.logger)
s6_conv3 = tf_utils.conv2d(x=s6_conv2,
output_dim=self.conv_dims[12],
k_h=3,
k_w=3,
d_h=1,
d_w=1,
padding='VALID',
initializer='He',
name='s6_conv3',
logger=self.logger)
s6_conv3 = tf_utils.relu(s6_conv3,
name='relu_s6_conv3',
logger=self.logger)
# Stage 7
s7_deconv1 = tf_utils.deconv2d(x=s6_conv3,
output_dim=self.conv_dims[13],
k_h=2,
k_w=2,
initializer='He',
name='s7_deconv1',
logger=self.logger)
s7_deconv1 = tf_utils.relu(s7_deconv1,
name='relu_s7_deconv1',
logger=self.logger)
# Cropping
h1, w1 = s3_conv2.get_shape().as_list()[1:3]
h2, w2 = s7_deconv1.get_shape().as_list()[1:3]
s3_conv2_crop = tf.image.crop_to_bounding_box(
image=s3_conv2,
offset_height=int(0.5 * (h1 - h2)),
offset_width=int(0.5 * (w1 - w2)),
target_height=h2,
target_width=w2)
tf_utils.print_activations(s3_conv2_crop, logger=self.logger)
s7_concat = tf_utils.concat(values=[s3_conv2_crop, s7_deconv1],
axis=3,
name='s7_concat',
logger=self.logger)
s7_conv2 = tf_utils.conv2d(x=s7_concat,
output_dim=self.conv_dims[14],
k_h=3,
k_w=3,
d_h=1,
d_w=1,
padding='VALID',
initializer='He',
name='s7_conv2',
logger=self.logger)
s7_conv2 = tf_utils.relu(s7_conv2,
name='relu_s7_conv2',
logger=self.logger)
s7_conv3 = tf_utils.conv2d(x=s7_conv2,
output_dim=self.conv_dims[15],
k_h=3,
k_w=3,
d_h=1,
d_w=1,
padding='VALID',
initializer='He',
name='s7_conv3',
logger=self.logger)
s7_conv3 = tf_utils.relu(s7_conv3,
name='relu_s7_conv3',
logger=self.logger)
# Stage 8
s8_deconv1 = tf_utils.deconv2d(x=s7_conv3,
output_dim=self.conv_dims[16],
k_h=2,
k_w=2,
initializer='He',
name='s8_deconv1',
logger=self.logger)
s8_deconv1 = tf_utils.relu(s8_deconv1,
name='relu_s8_deconv1',
logger=self.logger)
# Cropping
h1, w1 = s2_conv2.get_shape().as_list()[1:3]
h2, w2 = s8_deconv1.get_shape().as_list()[1:3]
s2_conv2_crop = tf.image.crop_to_bounding_box(
image=s2_conv2,
offset_height=int(0.5 * (h1 - h2)),
offset_width=int(0.5 * (w1 - w2)),
target_height=h2,
target_width=w2)
tf_utils.print_activations(s2_conv2_crop, logger=self.logger)
s8_concat = tf_utils.concat(values=[s2_conv2_crop, s8_deconv1],
axis=3,
name='s8_concat',
logger=self.logger)
s8_conv2 = tf_utils.conv2d(x=s8_concat,
output_dim=self.conv_dims[17],
k_h=3,
k_w=3,
d_h=1,
d_w=1,
padding='VALID',
initializer='He',
name='s8_conv2',
logger=self.logger)
s8_conv2 = tf_utils.relu(s8_conv2,
name='relu_s8_conv2',
logger=self.logger)
s8_conv3 = tf_utils.conv2d(x=s8_conv2,
output_dim=self.conv_dims[18],
k_h=3,
k_w=3,
d_h=1,
d_w=1,
padding='VALID',
initializer='He',
name='s8_conv3',
logger=self.logger)
s8_conv3 = tf_utils.relu(s8_conv3,
name='relu_conv3',
logger=self.logger)
# Stage 9
s9_deconv1 = tf_utils.deconv2d(x=s8_conv3,
output_dim=self.conv_dims[19],
k_h=2,
k_w=2,
initializer='He',
name='s9_deconv1',
logger=self.logger)
s9_deconv1 = tf_utils.relu(s9_deconv1,
name='relu_s9_deconv1',
logger=self.logger)
# Cropping
h1, w1 = s1_conv2.get_shape().as_list()[1:3]
h2, w2 = s9_deconv1.get_shape().as_list()[1:3]
s1_conv2_crop = tf.image.crop_to_bounding_box(
image=s1_conv2,
offset_height=int(0.5 * (h1 - h2)),
offset_width=int(0.5 * (w1 - w2)),
target_height=h2,
target_width=w2)
tf_utils.print_activations(s1_conv2_crop, logger=self.logger)
s9_concat = tf_utils.concat(values=[s1_conv2_crop, s9_deconv1],
axis=3,
name='s9_concat',
logger=self.logger)
s9_conv2 = tf_utils.conv2d(x=s9_concat,
output_dim=self.conv_dims[20],
k_h=3,
k_w=3,
d_h=1,
d_w=1,
padding='VALID',
initializer='He',
name='s9_conv2',
logger=self.logger)
s9_conv2 = tf_utils.relu(s9_conv2,
name='relu_s9_conv2',
logger=self.logger)
s9_conv3 = tf_utils.conv2d(x=s9_conv2,
output_dim=self.conv_dims[21],
k_h=3,
k_w=3,
d_h=1,
d_w=1,
padding='VALID',
initializer='He',
name='s9_conv3',
logger=self.logger)
s9_conv3 = tf_utils.relu(s9_conv3,
name='relu_s9_conv3',
logger=self.logger)
self.pred = tf_utils.conv2d(x=s9_conv3,
output_dim=self.conv_dims[22],
k_h=1,
k_w=1,
d_h=1,
d_w=1,
padding='SAME',
initializer='He',
name='output',
logger=self.logger)
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
