def layer_op(self, codes, is_training):
"""
:param codes: tensor, input latent space
:param is_training: boolean, True if network is in training mode
:return: samples from posterior distribution, means and log variances of the posterior distribution
"""
def clip(input):
# This is for clipping logvars,
# so that variances = exp(logvars) behaves well
output = tf.maximum(input, self.logvars_lower_bound)
output = tf.minimum(output, self.logvars_upper_bound)
return output
encoder_means = FullyConnectedLayer(
n_output_chns=self.number_of_latent_variables,
feature_normalization=None,
acti_func=None,
w_initializer=self.initializers['w'],
w_regularizer=self.regularizers['w'],
name='encoder_fc_means_{}'.format(self.number_of_latent_variables))
print(encoder_means)
encoder_logvars = FullyConnectedLayer(
n_output_chns=self.number_of_latent_variables,
feature_normalization=None,
acti_func=None,
w_initializer=self.initializers['w'],
w_regularizer=self.regularizers['w'],
name='encoder_fc_logvars_{}'.format(
self.number_of_latent_variables))
print(encoder_logvars)
# Predict the posterior distribution's parameters
posterior_means = encoder_means(codes, is_training)
posterior_logvars = clip(encoder_logvars(codes, is_training))
if self.number_of_samples == 1:
noise_sample = tf.random_normal(tf.shape(posterior_means), 0.0,
1.0)
else:
sample_shape = tf.concat([
tf.constant(self.number_of_samples, shape=[
1,
]),
tf.shape(posterior_means)
],
axis=0)
noise_sample = tf.reduce_mean(tf.random_normal(
sample_shape, 0.0, 1.0),
axis=0)
return [
posterior_means + tf.exp(0.5 * posterior_logvars) * noise_sample,
posterior_means, posterior_logvars
]
def layer_op(self, images, is_training, bn_momentum=0.9, layer_id=-1):
block1 = ConvPoolBlock((self.num_features[0], self.num_features[0]),
((1, 3, 3), (1, 3, 3)),
w_initializer=self.initializers['w'],
w_regularizer=self.regularizers['w'],
acti_func=self.acti_func,
name='B1')
block2 = ConvPoolBlock((self.num_features[1], self.num_features[1]),
((1, 3, 3), (1, 3, 3)),
w_initializer=self.initializers['w'],
w_regularizer=self.regularizers['w'],
acti_func=self.acti_func,
name='B2')
block3 = ConvPoolBlock((self.num_features[2], self.num_features[2]),
((1, 3, 3), (1, 3, 3), (1, 3, 3)),
w_initializer=self.initializers['w'],
w_regularizer=self.regularizers['w'],
acti_func=self.acti_func,
name='B3')
block4 = ConvPoolBlock((self.num_features[3], self.num_features[3]),
((1, 3, 3), (1, 3, 3)),
w_initializer=self.initializers['w'],
w_regularizer=self.regularizers['w'],
acti_func=self.acti_func,
name='B4')
block5 = ConvPoolBlock((self.num_features[4], self.num_features[4]),
((1, 3, 3), (1, 3, 3)),
w_initializer=self.initializers['w'],
w_regularizer=self.regularizers['w'],
acti_func=self.acti_func,
name='B5')
fc1 = FullyConnectedLayer(self.num_features[5],
w_initializer=self.initializers['w'],
w_regularizer=self.regularizers['w'],
acti_func=self.acti_func,
name='fc1')
fc2 = FullyConnectedLayer(self.num_out,
w_initializer=w_initializer_near_zero(),
w_regularizer=self.regularizers['w'],
with_bn=False,
acti_func=self.acti_func,
name='fc2')
output = block1(images, is_training, bn_momentum)
output = block2(output, is_training, bn_momentum)
output = block3(output, is_training, bn_momentum)
output = block4(output, is_training, bn_momentum)
output = block5(output, is_training, bn_momentum)
output = tf.reshape(output, [self.batch_size, -1])
output = fc1(output, is_training, bn_momentum)
output = fc2(output, is_training, bn_momentum)
return output
def layer_op(self, features):
batch_size = features.shape.as_list()[0]
conv_1 = ConvolutionalLayer(
20, 3, with_bn=False, with_bias=True, acti_func='relu')
fc_1 = FullyConnectedLayer(
20, with_bn=False, with_bias=True, acti_func='relu')
fc_2 = FullyConnectedLayer(
2, with_bn=False, with_bias=True)
hidden_feature = conv_1(features, is_training=True)
hidden_feature = tf.reshape(hidden_feature, [batch_size, -1])
hidden_feature = fc_1(hidden_feature, is_training=True)
logits = fc_2(hidden_feature, is_training=True)
return logits
def fully_connected(ch, features):
with tf.name_scope('fully_connected'):
# with bn?
fc_layer = FullyConnectedLayer(n_output_chns=ch,
with_bn=False,
with_bias=True)
return fc_layer(features, is_training=is_training)
def layer_op(self, codes, is_training):
"""
:param codes: tensor, input latent codes
:param is_training: boolean, True if network is in training mode
:return: tensor, output of series of FC layers
"""
# Define the decoding fully-connected layers
decoders_fc = []
for i in range(0, len(self.layer_sizes_decoder)):
decoders_fc.append(
FullyConnectedLayer(n_output_chns=self.layer_sizes_decoder[i],
with_bias=True,
feature_normalization='batch',
acti_func=self.acti_func_decoder[i],
w_initializer=self.initializers['w'],
w_regularizer=self.regularizers['w'],
name='FCDecoder_fc_{}'.format(
self.layer_sizes_decoder[i])))
print(decoders_fc[-1])
# Fully-connected decoder layers
flow = codes
for i in range(0, len(self.layer_sizes_decoder)):
flow = decoders_fc[i](flow, is_training)
return flow
def layer_op(self, input_tensor):
# spatial squeeze
input_rank = len(input_tensor.shape)
reduce_indices = list(range(input_rank))[1:-1]
if self.func == 'AVG':
squeeze_tensor = tf.reduce_mean(input_tensor, axis=reduce_indices)
elif self.func == 'MAX':
squeeze_tensor = tf.reduce_max(input_tensor, axis=reduce_indices)
else:
raise NotImplementedError("pooling function not supported")
# channel excitation
num_channels = int(squeeze_tensor.shape[-1])
reduction_ratio = self.reduction_ratio
if num_channels % reduction_ratio != 0:
raise ValueError(
"reduction ratio incompatible with "
"number of input tensor channels")
num_channels_reduced = num_channels / reduction_ratio
fc1 = FullyConnectedLayer(num_channels_reduced,
with_bias=False,
with_bn=False,
acti_func='relu',
name='se_fc_1')
fc2 = FullyConnectedLayer(num_channels,
with_bias=False,
with_bn=False,
acti_func='sigmoid',
name='se_fc_2')
fc_out_1 = fc1(squeeze_tensor)
fc_out_2 = fc2(fc_out_1)
while len(fc_out_2.shape) < input_rank:
fc_out_2 = tf.expand_dims(fc_out_2, axis=1)
output_tensor = tf.multiply(input_tensor, fc_out_2)
return output_tensor
def fully_connected(ch, features):
"""
Final discriminator processing block
:param ch: int, number of output channels
:param features: tensor, input features for final classification
:return: tensor, output logits of discriminator classification
"""
with tf.name_scope('fully_connected'):
# with bn?
fc_layer = FullyConnectedLayer(n_output_chns=ch,
feature_normalization=None,
with_bias=True)
return fc_layer(features, is_training=is_training)
def noise_to_image(sz, ch, rand_tensor, with_conditioning):
batch_size = rand_tensor.get_shape().as_list()[0]
output_shape = [batch_size] + sz + [ch]
with tf.name_scope('noise_to_image'):
g_no_0 = np.prod(sz) * ch
fc_layer = FullyConnectedLayer(
n_output_chns=g_no_0,
with_bn=False,
with_bias=True,
w_initializer=self.initializers['w'],
b_initializer=self.initializers['b'])
g_h1p = fc_layer(rand_tensor, keep_prob=keep_prob_ph)
g_h1p = tf.reshape(g_h1p, output_shape)
g_h1p = concat_cond(g_h1p, with_conditioning)
return conv(ch + conditioning_channels, g_h1p)
def _test_fc_layer_output_shape(self,
rank,
param_dict,
output_shape,
is_training=None,
dropout_prob=None):
if rank == 2:
input_data = self.get_2d_input()
elif rank == 3:
input_data = self.get_3d_input()
fc_layer = FullyConnectedLayer(**param_dict)
output_data = fc_layer(input_data,
is_training=is_training,
keep_prob=dropout_prob)
print(fc_layer)
with self.cached_session() as sess:
sess.run(tf.global_variables_initializer())
output_value = sess.run(output_data)
self.assertAllClose(output_shape, output_value.shape)
def layer_op(self, codes, is_training):
# Define the decoding fully-connected layers
decoders_fc = []
for i in range(0, len(self.layer_sizes_decoder)):
decoders_fc.append(FullyConnectedLayer(
n_output_chns=self.layer_sizes_decoder[i],
with_bias=True,
with_bn=True,
acti_func=self.acti_func_decoder[i],
w_initializer=self.initializers['w'],
w_regularizer=self.regularizers['w'],
name='FCDecoder_fc_{}'.format(self.layer_sizes_decoder[i])))
print(decoders_fc[-1])
# Fully-connected decoder layers
flow = codes
for i in range(0, len(self.layer_sizes_decoder)):
flow = decoders_fc[i](flow, is_training)
return flow
def noise_to_image(sz, ch, rand_tensor, with_conditioning):
"""
Processes random noise with fully connected layer and then convolutional layer
:param sz: image size
:param ch: int, number of output channels
:param rand_tensor: tensor, input random noise to generate image
:param with_conditioning: boolean, True if conditioning is to be used
:return: tensor, output of convolutional layer
"""
batch_size = rand_tensor.shape.as_list()[0]
output_shape = [batch_size] + sz + [ch]
with tf.name_scope('noise_to_image'):
g_no_0 = np.prod(sz) * ch
fc_layer = FullyConnectedLayer(
n_output_chns=g_no_0,
feature_normalization=None,
with_bias=True,
w_initializer=self.initializers['w'],
b_initializer=self.initializers['b'])
g_h1p = fc_layer(rand_tensor, keep_prob=keep_prob_ph)
g_h1p = tf.reshape(g_h1p, output_shape)
g_h1p = concat_cond(g_h1p, with_conditioning)
return conv(ch + conditioning_channels, g_h1p)
def layer_op(self, codes, is_training):
# Define the decoding fully-connected layers
decoders_fc = []
for i in range(0, len(self.layer_sizes_decoder)):
decoders_fc.append(FullyConnectedLayer(
n_output_chns=self.layer_sizes_decoder[i],
with_bias=True,
with_bn=True,
acti_func=self.acti_func_decoder[i],
w_initializer=self.initializers['w'],
w_regularizer=self.regularizers['w'],
name='decoder_fc_{}'.format(self.layer_sizes_decoder[i])))
print(decoders_fc[-1])
# Define the decoding convolutional layers
decoders_cnn = []
decoders_upsamplers = []
for i in range(0, len(self.trans_conv_output_channels)):
if self.upsampling_mode == 'DECONV':
decoders_upsamplers.append(DeconvolutionalLayer(
n_output_chns=self.trans_conv_output_channels[i],
kernel_size=self.trans_conv_unpooling_factors[i],
stride=self.trans_conv_unpooling_factors[i],
padding='SAME',
with_bias=True,
with_bn=True,
w_initializer=self.initializers['w'],
w_regularizer=None,
acti_func=None,
name='decoder_upsampler_{}_{}'.format(
self.trans_conv_unpooling_factors[i],
self.trans_conv_unpooling_factors[i])))
print(decoders_upsamplers[-1])
decoders_cnn.append(DeconvolutionalLayer(
n_output_chns=self.trans_conv_output_channels[i],
kernel_size=self.trans_conv_kernel_sizes[i],
stride=1,
padding='SAME',
with_bias=True,
with_bn=True,
#with_bn=not (i == len(self.trans_conv_output_channels) - 1),
# No BN on output
w_initializer=self.initializers['w'],
w_regularizer=None,
acti_func=self.acti_func_trans_conv[i],
name='decoder_trans_conv_{}_{}'.format(
self.trans_conv_kernel_sizes[i],
self.trans_conv_output_channels[i])))
print(decoders_cnn[-1])
# Fully-connected decoder layers
flow = codes
for i in range(0, len(self.layer_sizes_decoder)):
flow = decoders_fc[i](flow, is_training)
# Reconstitute the feature maps
flow = tf.reshape(flow, [-1] + self.downsampled_shape)
# Convolutional decoder layers
for i in range(0, len(self.trans_conv_output_channels)):
if self.upsampling_mode == 'DECONV':
flow = decoders_upsamplers[i](flow, is_training)
elif self.upsampling_mode == 'CHANNELWISE_DECONV':
flow = UpSampleLayer(
'CHANNELWISE_DECONV',
kernel_size=self.trans_conv_unpooling_factors[i],
stride=self.trans_conv_unpooling_factors[i])(flow)
elif self.upsampling_mode == 'REPLICATE':
flow = UpSampleLayer(
'REPLICATE',
kernel_size=self.trans_conv_unpooling_factors[i],
stride=self.trans_conv_unpooling_factors[i])(flow)
flow = decoders_cnn[i](flow, is_training)
return flow
def layer_op(self, images, is_training):
# Define the encoding convolutional layers
encoders_cnn = []
encoders_downsamplers = []
for i in range(0, len(self.conv_output_channels)):
encoders_cnn.append(ConvolutionalLayer(
n_output_chns=self.conv_output_channels[i],
kernel_size=self.conv_kernel_sizes[i],
padding='SAME',
with_bias=True,
with_bn=True,
w_initializer=self.initializers['w'],
w_regularizer=None,
acti_func=self.acti_func_conv[i],
name='encoder_conv_{}_{}'.format(
self.conv_kernel_sizes[i],
self.conv_output_channels[i])))
print(encoders_cnn[-1])
encoders_downsamplers.append(ConvolutionalLayer(
n_output_chns=self.conv_output_channels[i],
kernel_size=self.conv_pooling_factors[i],
stride=self.conv_pooling_factors[i],
padding='SAME',
with_bias=False,
with_bn=True,
w_initializer=self.initializers['w'],
w_regularizer=None,
acti_func=self.acti_func_conv[i],
name='encoder_downsampler_{}_{}'.format(
self.conv_pooling_factors[i],
self.conv_pooling_factors[i])))
print(encoders_downsamplers[-1])
# Define the encoding fully-connected layers
encoders_fc = []
for i in range(0, len(self.layer_sizes_encoder)):
encoders_fc.append(FullyConnectedLayer(
n_output_chns=self.layer_sizes_encoder[i],
with_bias=True,
with_bn=True,
acti_func=self.acti_func_encoder[i],
w_initializer=self.initializers['w'],
w_regularizer=self.regularizers['w'],
name='encoder_fc_{}'.format(self.layer_sizes_encoder[i])))
print(encoders_fc[-1])
# Add Gaussian noise to the input
if self.denoising_variance > 0 and is_training:
flow = images + tf.random_normal(
tf.shape(images), 0.0, self.denoising_variance)
else:
flow = images
# Convolutional encoder layers
for i in range(0, len(self.conv_output_channels)):
flow = encoders_downsamplers[i](
encoders_cnn[i](flow, is_training), is_training)
# Flatten the feature maps
flow = tf.reshape(flow, [-1, self.serialised_shape])
# Fully-connected encoder layers
for i in range(0, len(self.layer_sizes_encoder)):
flow = encoders_fc[i](flow, is_training)
return flow
def layer_op(self, images, is_training, keep_prob=None, num_units_max=None):
# Define the encoding convolutional layers
encoders_cnn = []
encoders_downsamplers = []
for i in range(0, len(self.conv_output_channels)):
encoders_cnn.append(ConvolutionalLayer(
n_output_chns=self.conv_output_channels[i],
kernel_size=self.conv_kernel_sizes[i],
stride=self.conv_kernel_stride[i],
padding='SAME',
with_bias=True,
feature_normalization='batch',
w_initializer=self.initializers['w'],
w_regularizer=None,
acti_func=self.acti_func_conv[i],
name='encoder_conv_{}_{}'.format(
self.conv_kernel_sizes[i],
self.conv_output_channels[i])))
print(encoders_cnn[-1])
encoders_downsamplers.append(ConvolutionalLayer(
n_output_chns=self.conv_output_channels[i],
kernel_size=self.conv_pooling_factors[i],
stride=self.conv_pooling_factors[i],
padding='SAME',
with_bias=False,
feature_normalization='batch',
w_initializer=self.initializers['w'],
w_regularizer=None,
acti_func=self.acti_func_conv[i],
name='encoder_downsampler_{}_{}'.format(
self.conv_pooling_factors[i],
self.conv_pooling_factors[i])))
print(encoders_downsamplers[-1])
# Define the encoding fully-connected layers
encoders_fc = []
for i in range(0, len(self.layer_sizes_encoder)):
encoders_fc.append(FullyConnectedLayer(
n_output_chns=self.layer_sizes_encoder[i],
with_bias=True,
feature_normalization='batch',
acti_func=self.acti_func_encoder[i],
w_initializer=self.initializers['w'],
w_regularizer=self.regularizers['w'],
name='encoder_fc_{}'.format(self.layer_sizes_encoder[i])))
print(encoders_fc[-1])
# Add Gaussian noise to the input
if self.denoising_variance > 0 and is_training:
flow = images + tf.random_normal(
tf.shape(images), 0.0, self.denoising_variance)
else:
flow = images
# Convolutional encoder layers
print("num_units_max: ", num_units_max)
for i in range(0, len(self.conv_output_channels)):
if keep_prob is not None:
if type(keep_prob) == float: # if the same probability of dropout on all the layers
flow = encoders_downsamplers[i](
encoders_cnn[i](flow, is_training, keep_prob=keep_prob), is_training)
else: # if keep_prob is a list
flow = encoders_downsamplers[i](
encoders_cnn[i](flow, is_training, keep_prob=keep_prob[i]), is_training)
elif num_units_max is not None:
if (type(num_units_max) == float) | (type(num_units_max) == int): # if the same probability of dropout on all the layers
flow = encoders_downsamplers[i](
encoders_cnn[i](flow, is_training, num_units_max=num_units_max), is_training)
else: # if keep_prob is a list
flow = encoders_downsamplers[i](
encoders_cnn[i](flow, is_training, num_units_max=num_units_max[i]), is_training)
else: # if no dropout nor maxout
flow = encoders_downsamplers[i](
encoders_cnn[i](flow, is_training), is_training)
print("Conv block ", i, ": ", flow.shape)
# Flatten the feature maps
#if len(self.layer_sizes_encoder) > 0:
# flow = tf.reshape(flow, [-1, self.serialised_shape])
print("After convolution, flow shape: ", flow.shape)
# Fully-connected encoder layers
for i in range(0, len(self.layer_sizes_encoder)):
flow = encoders_fc[i](flow, is_training)
print("FC ", i, ": ", flow.shape)
return flow