def inference(self, input_layer, reuse, is_training, name, print_shape=True):
"""Forward propagation of computational graph."""
# Check input layer dimensions
assert input_layer.shape[1] == self.length
assert input_layer.shape[2] == self.channels
# Define a scope for reusing the variables
with tf.variable_scope(name, reuse=reuse):
# Set variables
kernel_size = 3
conv_filts = 128
res_filts = 128
skip_filts = 128
skips = list()
# Print shape
print_output_shape(layer_name='input', net=input_layer, print_shape=print_shape)
"""Block Series 1"""
# --- Layer 1 (Convolution) ------------------------------------------------------------------------------ #
# Set name
layer_name = 'layer_1'
# Set layer scope
with tf.variable_scope(layer_name):
# Convolution
net = conv_layer(input_layer=input_layer, kernel_size=kernel_size, strides=1, dilation_rate=1,
filters=res_filts, padding='SAME', activation=None, use_bias=False,
name=layer_name + '_conv', seed=self.seed)
# Print shape
print_output_shape(layer_name=layer_name, net=net, print_shape=print_shape)
# --- Layer 2 (Convolution) ------------------------------------------------------------------------------ #
# Set name
layer_name = 'layer_2'
# Compute block
outputs = self._residual_block(input_layer=net, kernel_size=kernel_size, layer_name=layer_name,
conv_filts=conv_filts, res_filts=res_filts, skip_filts=skip_filts,
dilation_rate=2, res=True, skip=True)
# Collect skip
skips.append(outputs['skip'])
# Print shape
print_output_shape(layer_name=layer_name + '_res', net=outputs['res'], print_shape=print_shape)
print_output_shape(layer_name=layer_name + '_skip', net=outputs['skip'], print_shape=print_shape)
# --- Layer 3 (Convolution) ------------------------------------------------------------------------------ #
# Set name
layer_name = 'layer_3'
# Compute block
outputs = self._residual_block(input_layer=outputs['res'], kernel_size=kernel_size, layer_name=layer_name,
conv_filts=conv_filts, res_filts=res_filts, skip_filts=skip_filts,
dilation_rate=4, res=True, skip=True)
# Collect skip
skips.append(outputs['skip'])
# Print shape
print_output_shape(layer_name=layer_name + '_res', net=outputs['res'], print_shape=print_shape)
print_output_shape(layer_name=layer_name + '_skip', net=outputs['skip'], print_shape=print_shape)
# --- Layer 4 (Convolution) ------------------------------------------------------------------------------ #
# Set name
layer_name = 'layer_4'
# Compute block
outputs = self._residual_block(input_layer=outputs['res'], kernel_size=kernel_size, layer_name=layer_name,
conv_filts=conv_filts, res_filts=res_filts, skip_filts=skip_filts,
dilation_rate=8, res=True, skip=True)
# Collect skip
skips.append(outputs['skip'])
# Print shape
print_output_shape(layer_name=layer_name + '_res', net=outputs['res'], print_shape=print_shape)
print_output_shape(layer_name=layer_name + '_skip', net=outputs['skip'], print_shape=print_shape)
# --- Layer 5 (Convolution) ------------------------------------------------------------------------------ #
# Set name
layer_name = 'layer_5'
# Compute block
outputs = self._residual_block(input_layer=outputs['res'], kernel_size=kernel_size, layer_name=layer_name,
conv_filts=conv_filts, res_filts=res_filts, skip_filts=skip_filts,
dilation_rate=16, res=True, skip=True)
# Collect skip
skips.append(outputs['skip'])
# Print shape
print_output_shape(layer_name=layer_name + '_res', net=outputs['res'], print_shape=print_shape)
print_output_shape(layer_name=layer_name + '_skip', net=outputs['skip'], print_shape=print_shape)
# --- Layer 6 (Convolution) ------------------------------------------------------------------------------ #
# Set name
layer_name = 'layer_6'
# Compute block
outputs = self._residual_block(input_layer=outputs['res'], kernel_size=kernel_size, layer_name=layer_name,
conv_filts=conv_filts, res_filts=res_filts, skip_filts=skip_filts,
dilation_rate=32, res=True, skip=True)
# Collect skip
skips.append(outputs['skip'])
# Print shape
print_output_shape(layer_name=layer_name + '_res', net=outputs['res'], print_shape=print_shape)
print_output_shape(layer_name=layer_name + '_skip', net=outputs['skip'], print_shape=print_shape)
# --- Layer 7 (Convolution) ------------------------------------------------------------------------------ #
# Set name
layer_name = 'layer_7'
# Compute block
outputs = self._residual_block(input_layer=outputs['res'], kernel_size=kernel_size, layer_name=layer_name,
conv_filts=conv_filts, res_filts=res_filts, skip_filts=skip_filts,
dilation_rate=64, res=True, skip=True)
# Collect skip
skips.append(outputs['skip'])
# Print shape
print_output_shape(layer_name=layer_name + '_res', net=outputs['res'], print_shape=print_shape)
print_output_shape(layer_name=layer_name + '_skip', net=outputs['skip'], print_shape=print_shape)
# --- Layer 8 (Convolution) ------------------------------------------------------------------------------ #
# Set name
layer_name = 'layer_8'
# Compute block
outputs = self._residual_block(input_layer=outputs['res'], kernel_size=kernel_size, layer_name=layer_name,
conv_filts=conv_filts, res_filts=res_filts, skip_filts=skip_filts,
dilation_rate=128, res=True, skip=True)
# Collect skip
skips.append(outputs['skip'])
# Print shape
print_output_shape(layer_name=layer_name + '_res', net=outputs['res'], print_shape=print_shape)
print_output_shape(layer_name=layer_name + '_skip', net=outputs['skip'], print_shape=print_shape)
# --- Layer 9 (Convolution) ------------------------------------------------------------------------------ #
# Set name
layer_name = 'layer_9'
# Compute block
outputs = self._residual_block(input_layer=outputs['res'], kernel_size=kernel_size, layer_name=layer_name,
conv_filts=conv_filts, res_filts=res_filts, skip_filts=skip_filts,
dilation_rate=256, res=True, skip=True)
# Collect skip
skips.append(outputs['skip'])
# Print shape
print_output_shape(layer_name=layer_name + '_res', net=outputs['res'], print_shape=print_shape)
print_output_shape(layer_name=layer_name + '_skip', net=outputs['skip'], print_shape=print_shape)
# --- Layer 10 (Convolution) ----------------------------------------------------------------------------- #
# Set name
layer_name = 'layer_10'
# Compute block
outputs = self._residual_block(input_layer=outputs['res'], kernel_size=kernel_size, layer_name=layer_name,
conv_filts=conv_filts, res_filts=res_filts, skip_filts=skip_filts,
dilation_rate=512, res=False, skip=True)
# Collect skip
skips.append(outputs['skip'])
# Print shape
print_output_shape(layer_name=layer_name + '_skip', net=outputs['skip'], print_shape=print_shape)
# Add all skips to res output
with tf.variable_scope('skips'):
output = tf.add_n(inputs=skips, name='add_skips')
# Print shape
print_output_shape(layer_name='output_skip_addition', net=output, print_shape=print_shape)
# Activation
with tf.variable_scope('relu') as scope:
output = tf.nn.relu(output, name=scope.name)
# Dropout
output = dropout_layer(input_layer=output, drop_rate=0.3, seed=self.seed,
training=is_training, name='dropout1')
# Convolution
output = conv_layer(input_layer=output, kernel_size=kernel_size, strides=1, dilation_rate=1,
filters=256, padding='SAME', activation=tf.nn.relu, use_bias=False,
name='conv1', seed=self.seed)
# Dropout
output = dropout_layer(input_layer=output, drop_rate=0.3, seed=self.seed,
training=is_training, name='dropout1')
# Print shape
print_output_shape(layer_name='output_conv1', net=output, print_shape=print_shape)
# Convolution
output = conv_layer(input_layer=output, kernel_size=kernel_size, strides=1, dilation_rate=1,
filters=512, padding='SAME', activation=tf.nn.relu, use_bias=False,
name='conv2', seed=self.seed)
# Dropout
output = dropout_layer(input_layer=output, drop_rate=0.3, seed=self.seed,
training=is_training, name='dropout1')
# Print shape
print_output_shape(layer_name='output_conv2', net=output, print_shape=print_shape)
"""Network Output"""
# --- Global Average Pooling Layer ----------------------------------------------------------------------- #
# Set name
layer_name = 'gap'
# Set layer scope
with tf.variable_scope(layer_name):
# Reduce mean along dimension 1
gap = tf.reduce_mean(input_tensor=output, axis=1)
# Print shape
print_output_shape(layer_name=layer_name, net=gap, print_shape=print_shape)
# --- Softmax Layer -------------------------------------------------------------------------------------- #
# Set name
layer_name = 'logits'
# Softmax activation
logits = fc_layer(input_layer=gap, neurons=self.classes, activation=None, use_bias=False,
name=layer_name, seed=self.seed)
# Print shape
print_output_shape(layer_name=layer_name, net=logits, print_shape=print_shape)
# Compute Class Activation Maps
cams = self._get_cams(net=output, is_training=is_training)
return logits, cams
def inference(self,
input_layer,
reuse,
is_training,
name,
print_shape=True):
"""Forward propagation of computational graph."""
# Check input layer dimensions
assert input_layer.shape[1] == self.length
assert input_layer.shape[2] == self.channels
# Define a scope for reusing the variables
with tf.variable_scope(name, reuse=reuse):
"""Network Stem"""
# --- Layer 1 (Convolution) ------------------------------------------------------------------------------ #
# Set name
layer_name = 'layer_1'
# Set layer scope
with tf.variable_scope(layer_name):
# Convolution
net = conv_layer(input_layer=input_layer,
kernel_size=3,
strides=2,
dilation_rate=1,
filters=32,
padding='VALID',
activation=tf.nn.relu,
use_bias=True,
name=layer_name + '_conv_ks3_dr1',
seed=self.seed)
# Batch Norm
net = batch_norm_layer(input_layer=net,
training=is_training,
name=layer_name + '_batchnorm')
# Dropout
net = dropout_layer(input_layer=net,
drop_rate=0.3,
seed=self.seed,
training=is_training,
name=layer_name + '_dropout')
# Print shape
print_output_shape(layer_name=layer_name,
net=net,
print_shape=print_shape)
# --- Layer 2 (Convolution) ------------------------------------------------------------------------------ #
# Set name
layer_name = 'layer_2'
# Set layer scope
with tf.variable_scope(layer_name):
# Convolution
net = conv_layer(input_layer=net,
kernel_size=3,
strides=1,
dilation_rate=1,
filters=32,
padding='VALID',
activation=tf.nn.relu,
use_bias=True,
name=layer_name + '_conv_ks3_dr1',
seed=self.seed)
# Batch Norm
net = batch_norm_layer(input_layer=net,
training=is_training,
name=layer_name + '_batchnorm')
# Dropout
net = dropout_layer(input_layer=net,
drop_rate=0.3,
seed=self.seed,
training=is_training,
name=layer_name + '_dropout')
# Print shape
print_output_shape(layer_name=layer_name,
net=net,
print_shape=print_shape)
# --- Layer 3 (Convolution) ------------------------------------------------------------------------------ #
# Set name
layer_name = 'layer_3'
# Set layer scope
with tf.variable_scope(layer_name):
# Convolution
net = conv_layer(input_layer=net,
kernel_size=3,
strides=1,
dilation_rate=1,
filters=64,
padding='VALID',
activation=tf.nn.relu,
use_bias=True,
name=layer_name + '_conv_ks3_dr1',
seed=self.seed)
# Batch Norm
net = batch_norm_layer(input_layer=net,
training=is_training,
name=layer_name + '_batchnorm')
# Dropout
net = dropout_layer(input_layer=net,
drop_rate=0.3,
seed=self.seed,
training=is_training,
name=layer_name + '_dropout')
# Print shape
print_output_shape(layer_name=layer_name,
net=net,
print_shape=print_shape)
# --- Layer 4 (Mixed - Convolution) ---------------------------------------------------------------------- #
# Set name
layer_name = 'layer_4'
# Set layer scope
with tf.variable_scope(layer_name):
# Branch 0
with tf.variable_scope('branch_0'):
# Max pool
branch_0 = max_pool_layer(input_layer=net,
pool_size=3,
strides=2,
padding='VALID',
name=layer_name +
'_branch_0_a_maxpool_ps3')
# Branch 1
with tf.variable_scope('branch_1'):
# Convolution
branch_1 = conv_layer(input_layer=net,
kernel_size=3,
strides=2,
dilation_rate=1,
filters=96,
padding='VALID',
activation=tf.nn.relu,
use_bias=True,
name=layer_name +
'_branch_1_a_conv_ks3_dr1',
seed=self.seed)
# Merge branches
net = tf.concat(values=[branch_0, branch_1], axis=2)
# Batch Norm
net = batch_norm_layer(input_layer=net,
training=is_training,
name=layer_name + '_batchnorm')
# Dropout
net = dropout_layer(input_layer=net,
drop_rate=0.3,
seed=self.seed,
training=is_training,
name=layer_name + '_dropout')
# Print shape
print_output_shape(layer_name=layer_name,
net=net,
print_shape=print_shape)
# --- Layer 5 (Mixed - Convolution) ---------------------------------------------------------------------- #
# Set name
layer_name = 'layer_5'
# Set layer scope
with tf.variable_scope(layer_name):
# Branch 0
with tf.variable_scope('branch_0'):
# Convolution
branch_0 = conv_layer(input_layer=net,
kernel_size=1,
strides=1,
dilation_rate=1,
filters=64,
padding='SAME',
activation=tf.nn.relu,
use_bias=True,
name=layer_name +
'_branch_0_a_conv_ks1_dr1',
seed=self.seed)
branch_0 = conv_layer(input_layer=branch_0,
kernel_size=3,
strides=1,
dilation_rate=1,
filters=96,
padding='VALID',
activation=tf.nn.relu,
use_bias=True,
name=layer_name +
'_branch_0_b_conv_ks3_dr1',
seed=self.seed)
# Branch 1
with tf.variable_scope('branch_1'):
# Convolution
branch_1 = conv_layer(input_layer=net,
kernel_size=1,
strides=1,
dilation_rate=1,
filters=64,
padding='SAME',
activation=tf.nn.relu,
use_bias=True,
name=layer_name +
'_branch_1_a_conv_ks1_dr1',
seed=self.seed)
branch_1 = conv_layer(input_layer=branch_1,
kernel_size=7,
strides=1,
dilation_rate=2,
filters=64,
padding='SAME',
activation=tf.nn.relu,
use_bias=True,
name=layer_name +
'_branch_1_b_conv_ks7_dr2',
seed=self.seed)
branch_1 = conv_layer(input_layer=branch_1,
kernel_size=7,
strides=1,
dilation_rate=4,
filters=64,
padding='SAME',
activation=tf.nn.relu,
use_bias=True,
name=layer_name +
'_branch_1_c_conv_ks7_dr4',
seed=self.seed)
branch_1 = conv_layer(input_layer=branch_1,
kernel_size=3,
strides=1,
dilation_rate=1,
filters=96,
padding='VALID',
activation=tf.nn.relu,
use_bias=True,
name=layer_name +
'_branch_1_d_conv_ks3_dr1',
seed=self.seed)
# Merge branches
net = tf.concat(values=[branch_0, branch_1], axis=2)
# Batch Norm
net = batch_norm_layer(input_layer=net,
training=is_training,
name=layer_name + '_batchnorm')
# Dropout
net = dropout_layer(input_layer=net,
drop_rate=0.3,
seed=self.seed,
training=is_training,
name=layer_name + '_dropout')
# Print shape
print_output_shape(layer_name=layer_name,
net=net,
print_shape=print_shape)
# --- Layer 6 (Mixed - Convolution) ---------------------------------------------------------------------- #
# Set name
layer_name = 'layer_6'
# Set layer scope
with tf.variable_scope(layer_name):
# Branch 0
with tf.variable_scope('branch_0'):
# Convolution
branch_0 = conv_layer(input_layer=net,
kernel_size=3,
strides=2,
dilation_rate=1,
filters=192,
padding='VALID',
activation=tf.nn.relu,
use_bias=True,
name=layer_name +
'_branch_0_a_conv_ks3_dr1',
seed=self.seed)
# Branch 1
with tf.variable_scope('branch_1'):
# Max pool
branch_1 = max_pool_layer(input_layer=net,
pool_size=3,
strides=2,
padding='VALID',
name=layer_name +
'_branch_0_a_maxpool_ps3')
# Merge branches
net = tf.concat(values=[branch_0, branch_1], axis=2)
# Batch Norm
net = batch_norm_layer(input_layer=net,
training=is_training,
name=layer_name + '_batchnorm')
# Dropout
net = dropout_layer(input_layer=net,
drop_rate=0.3,
seed=self.seed,
training=is_training,
name=layer_name + '_dropout')
# Print shape
print_output_shape(layer_name=layer_name,
net=net,
print_shape=print_shape)
"""Inception-A (V4)"""
# Set number of Inception layers and previous layer
num_layers = 2
previous_layer = 6
for layer_id in np.arange(num_layers) + (previous_layer + 1):
# Set name
layer_name = 'layer_{}'.format(layer_id)
# Set layer scope
with tf.variable_scope(layer_name):
# Inception-A (V4)
net = self._inception_a(inputs=net, layer_name=layer_name)
# Batch Norm
net = batch_norm_layer(input_layer=net,
training=is_training,
name=layer_name + '_batchnorm')
# Print shape
print_output_shape(layer_name=layer_name,
net=net,
print_shape=print_shape)
"""Reduction-A"""
"""Inception-B"""
"""Reduction-B"""
"""Inception-C"""
"""Network Output"""
# --- Global Average Pooling Layer ----------------------------------------------------------------------- #
# Set layer scope
with tf.variable_scope('GAP'):
# Reduce mean along dimension 1
gap = tf.reduce_mean(input_tensor=net, axis=1)
# --- Softmax Layer -------------------------------------------------------------------------------------- #
# Softmax activation
logits = fc_layer(input_layer=gap,
neurons=self.classes,
activation=None,
use_bias=False,
name='logits',
seed=self.seed)
return logits, net
def inference(self,
input_layer,
reuse,
is_training,
name,
print_shape=True):
"""Forward propagation of computational graph."""
# Check input layer dimensions
assert input_layer.shape[1] == self.length
assert input_layer.shape[2] == self.channels
# Define a scope for reusing the variables
with tf.variable_scope(name, reuse=reuse):
"""Block Series 1"""
# --- Layer 1 (Convolution) ------------------------------------------------------------------------------ #
# Set name
layer_name = 'layer_1'
# Set layer scope
with tf.variable_scope(layer_name):
# Convolution
net = conv_layer(input_layer=input_layer,
kernel_size=3,
strides=1,
dilation_rate=1,
filters=128,
padding='SAME',
activation=None,
use_bias=False,
name=layer_name + '_conv_ks3_dr1',
seed=self.seed)
# Max pool
net = max_pool_layer(input_layer=net,
pool_size=3,
strides=2,
padding='SAME',
name=layer_name + '_maxpool_ps3')
# Print shape
print_output_shape(layer_name=layer_name,
net=net,
print_shape=print_shape)
# --- Layer 2 (Convolution) ------------------------------------------------------------------------------ #
# Set name
layer_name = 'layer_2'
# Set identity
identity = tf.identity(input=net, name=layer_name + '_identity')
# Set layer scope
with tf.variable_scope(layer_name):
# Convolution
net = conv_layer(input_layer=net,
kernel_size=3,
strides=1,
dilation_rate=2,
filters=128,
padding='SAME',
activation=None,
use_bias=False,
name=layer_name + '_conv_ks3_dr2',
seed=self.seed)
# Tanh activation
tanh = tf.nn.tanh(net, name='tanh')
# Sigmoid activation
sigmoid = tf.nn.sigmoid(net, name='sigmoid')
# Combine activations
net = tf.multiply(tanh, sigmoid, name='activation_gate')
# Convolution
net = conv_layer(input_layer=net,
kernel_size=1,
strides=1,
dilation_rate=1,
filters=128,
padding='SAME',
activation=None,
use_bias=False,
name=layer_name + '_conv_ks1_dr1',
seed=self.seed)
# Add identity
net = tf.add(net, identity, name=layer_name + '_identity_add')
# Print shape
print_output_shape(layer_name=layer_name,
net=net,
print_shape=print_shape)
# --- Layer 3 (Convolution) ------------------------------------------------------------------------------ #
# Set name
layer_name = 'layer_3'
# Set identity
identity = tf.identity(input=net, name=layer_name + '_identity')
# Set layer scope
with tf.variable_scope(layer_name):
# Convolution
net = conv_layer(input_layer=net,
kernel_size=3,
strides=1,
dilation_rate=4,
filters=128,
padding='SAME',
activation=None,
use_bias=False,
name=layer_name + '_conv_ks3_dr4',
seed=self.seed)
# Tanh activation
tanh = tf.nn.tanh(net, name='tanh')
# Sigmoid activation
sigmoid = tf.nn.sigmoid(net, name='sigmoid')
# Combine activations
net = tf.multiply(tanh, sigmoid, name='activation_gate')
# Convolution
net = conv_layer(input_layer=net,
kernel_size=1,
strides=1,
dilation_rate=1,
filters=128,
padding='SAME',
activation=None,
use_bias=False,
name=layer_name + '_conv_ks1_dr1',
seed=self.seed)
# Add identity
net = tf.add(net, identity, name=layer_name + '_identity_add')
# Print shape
print_output_shape(layer_name=layer_name,
net=net,
print_shape=print_shape)
# --- Layer 4 (Convolution) ------------------------------------------------------------------------------ #
# Set name
layer_name = 'layer_4'
# Set identity
identity = tf.identity(input=net, name=layer_name + '_identity')
# Set layer scope
with tf.variable_scope(layer_name):
# Convolution
net = conv_layer(input_layer=net,
kernel_size=3,
strides=1,
dilation_rate=8,
filters=128,
padding='SAME',
activation=None,
use_bias=False,
name=layer_name + '_conv_ks3_dr8',
seed=self.seed)
# Tanh activation
tanh = tf.nn.tanh(net, name='tanh')
# Sigmoid activation
sigmoid = tf.nn.sigmoid(net, name='sigmoid')
# Combine activations
net = tf.multiply(tanh, sigmoid, name='activation_gate')
# Convolution
net = conv_layer(input_layer=net,
kernel_size=1,
strides=1,
dilation_rate=1,
filters=128,
padding='SAME',
activation=None,
use_bias=False,
name=layer_name + '_conv_ks1_dr1',
seed=self.seed)
# Add identity
net = tf.add(net, identity, name=layer_name + '_identity_add')
# Print shape
print_output_shape(layer_name=layer_name,
net=net,
print_shape=print_shape)
# --- Layer 5 (Convolution) ------------------------------------------------------------------------------ #
# Set name
layer_name = 'layer_5'
# Set identity
identity = tf.identity(input=net, name=layer_name + '_identity')
# Set layer scope
with tf.variable_scope(layer_name):
# Convolution
net = conv_layer(input_layer=net,
kernel_size=3,
strides=1,
dilation_rate=16,
filters=128,
padding='SAME',
activation=None,
use_bias=False,
name=layer_name + '_conv_ks3_dr16',
seed=self.seed)
# Tanh activation
tanh = tf.nn.tanh(net, name='tanh')
# Sigmoid activation
sigmoid = tf.nn.sigmoid(net, name='sigmoid')
# Combine activations
net = tf.multiply(tanh, sigmoid, name='activation_gate')
# Convolution
net = conv_layer(input_layer=net,
kernel_size=1,
strides=1,
dilation_rate=1,
filters=128,
padding='SAME',
activation=None,
use_bias=False,
name=layer_name + '_conv_ks1_dr1',
seed=self.seed)
# Add identity
net = tf.add(net, identity, name=layer_name + '_identity_add')
# Print shape
print_output_shape(layer_name=layer_name,
net=net,
print_shape=print_shape)
# --- Layer 6 (Convolution) ------------------------------------------------------------------------------ #
# Set name
layer_name = 'layer_6'
# Set identity
identity = tf.identity(input=net, name=layer_name + '_identity')
# Set layer scope
with tf.variable_scope(layer_name):
# Convolution
net = conv_layer(input_layer=net,
kernel_size=3,
strides=1,
dilation_rate=32,
filters=128,
padding='SAME',
activation=None,
use_bias=False,
name=layer_name + '_conv_ks3_dr32',
seed=self.seed)
# Tanh activation
tanh = tf.nn.tanh(net, name='tanh')
# Sigmoid activation
sigmoid = tf.nn.sigmoid(net, name='sigmoid')
# Combine activations
net = tf.multiply(tanh, sigmoid, name='activation_gate')
# Convolution
net = conv_layer(input_layer=net,
kernel_size=1,
strides=1,
dilation_rate=1,
filters=128,
padding='SAME',
activation=None,
use_bias=False,
name=layer_name + '_conv_ks1_dr1',
seed=self.seed)
# Add identity
net = tf.add(net, identity, name=layer_name + '_identity_add')
# Print shape
print_output_shape(layer_name=layer_name,
net=net,
print_shape=print_shape)
# --- Layer 7 (Convolution) ------------------------------------------------------------------------------ #
# Set name
layer_name = 'layer_7'
# Set identity
identity = tf.identity(input=net, name=layer_name + '_identity')
# Set layer scope
with tf.variable_scope(layer_name):
# Convolution
net = conv_layer(input_layer=net,
kernel_size=3,
strides=1,
dilation_rate=64,
filters=128,
padding='SAME',
activation=None,
use_bias=False,
name=layer_name + '_conv_ks3_dr64',
seed=self.seed)
# Tanh activation
tanh = tf.nn.tanh(net, name='tanh')
# Sigmoid activation
sigmoid = tf.nn.sigmoid(net, name='sigmoid')
# Combine activations
net = tf.multiply(tanh, sigmoid, name='activation_gate')
# Convolution
net = conv_layer(input_layer=net,
kernel_size=1,
strides=1,
dilation_rate=1,
filters=128,
padding='SAME',
activation=None,
use_bias=False,
name=layer_name + '_conv_ks1_dr1',
seed=self.seed)
# Add identity
net = tf.add(net, identity, name=layer_name + '_identity_add')
# Print shape
print_output_shape(layer_name=layer_name,
net=net,
print_shape=print_shape)
# --- Layer 8 (Convolution) ------------------------------------------------------------------------------ #
# Set name
layer_name = 'layer_8'
# Set identity
identity = tf.identity(input=net, name=layer_name + '_identity')
# Set layer scope
with tf.variable_scope(layer_name):
# Convolution
net = conv_layer(input_layer=net,
kernel_size=3,
strides=1,
dilation_rate=128,
filters=128,
padding='SAME',
activation=None,
use_bias=False,
name=layer_name + '_conv_ks3_dr128',
seed=self.seed)
# Tanh activation
tanh = tf.nn.tanh(net, name='tanh')
# Sigmoid activation
sigmoid = tf.nn.sigmoid(net, name='sigmoid')
# Combine activations
net = tf.multiply(tanh, sigmoid, name='activation_gate')
# Convolution
net = conv_layer(input_layer=net,
kernel_size=1,
strides=1,
dilation_rate=1,
filters=128,
padding='SAME',
activation=None,
use_bias=False,
name=layer_name + '_conv_ks1_dr1',
seed=self.seed)
# Add identity
net = tf.add(net, identity, name=layer_name + '_identity_add')
# Print shape
print_output_shape(layer_name=layer_name,
net=net,
print_shape=print_shape)
# --- Layer 9 (Convolution) ------------------------------------------------------------------------------ #
# Set name
layer_name = 'layer_9'
# Set identity
identity = tf.identity(input=net, name=layer_name + '_identity')
# Set layer scope
with tf.variable_scope(layer_name):
# Convolution
net = conv_layer(input_layer=net,
kernel_size=3,
strides=1,
dilation_rate=256,
filters=128,
padding='SAME',
activation=None,
use_bias=False,
name=layer_name + '_conv_ks3_dr256',
seed=self.seed)
# Tanh activation
tanh = tf.nn.tanh(net, name='tanh')
# Sigmoid activation
sigmoid = tf.nn.sigmoid(net, name='sigmoid')
# Combine activations
net = tf.multiply(tanh, sigmoid, name='activation_gate')
# Convolution
net = conv_layer(input_layer=net,
kernel_size=1,
strides=1,
dilation_rate=1,
filters=128,
padding='SAME',
activation=None,
use_bias=False,
name=layer_name + '_conv_ks1_dr1',
seed=self.seed)
# Add identity
net = tf.add(net, identity, name=layer_name + '_identity_add')
# Print shape
print_output_shape(layer_name=layer_name,
net=net,
print_shape=print_shape)
# --- Layer 10 (Convolution) ----------------------------------------------------------------------------- #
# Set name
layer_name = 'layer_10'
# Set identity
identity = tf.identity(input=net, name=layer_name + '_identity')
# Set layer scope
with tf.variable_scope(layer_name):
# Convolution
net = conv_layer(input_layer=net,
kernel_size=3,
strides=1,
dilation_rate=512,
filters=128,
padding='SAME',
activation=None,
use_bias=False,
name=layer_name + '_conv_ks3_dr512',
seed=self.seed)
# Tanh activation
tanh = tf.nn.tanh(net, name='tanh')
# Sigmoid activation
sigmoid = tf.nn.sigmoid(net, name='sigmoid')
# Combine activations
net = tf.multiply(tanh, sigmoid, name='activation_gate')
# Convolution
net = conv_layer(input_layer=net,
kernel_size=1,
strides=1,
dilation_rate=1,
filters=128,
padding='SAME',
activation=None,
use_bias=False,
name=layer_name + '_conv_ks1_dr1',
seed=self.seed)
# Add identity
net = tf.add(net, identity, name=layer_name + '_identity_add')
# Print shape
print_output_shape(layer_name=layer_name,
net=net,
print_shape=print_shape)
# --- Layer 11 (Convolution) ----------------------------------------------------------------------------- #
# Set name
layer_name = 'layer_11'
# Set identity
identity = tf.identity(input=net, name=layer_name + '_identity')
# Set layer scope
with tf.variable_scope(layer_name):
# Convolution
net = conv_layer(input_layer=net,
kernel_size=3,
strides=1,
dilation_rate=1,
filters=128,
padding='SAME',
activation=None,
use_bias=False,
name=layer_name + '_conv_ks3_dr128',
seed=self.seed)
# Tanh activation
tanh = tf.nn.tanh(net, name='tanh')
# Sigmoid activation
sigmoid = tf.nn.sigmoid(net, name='sigmoid')
# Combine activations
net = tf.multiply(tanh, sigmoid, name='activation_gate')
# Convolution
net = conv_layer(input_layer=net,
kernel_size=1,
strides=1,
dilation_rate=1,
filters=128,
padding='SAME',
activation=None,
use_bias=False,
name=layer_name + '_conv_ks1_dr1',
seed=self.seed)
# Add identity
net = tf.add(net, identity, name=layer_name + '_identity_add')
# Print shape
print_output_shape(layer_name=layer_name,
net=net,
print_shape=print_shape)
# --- Layer 12 (Convolution) ----------------------------------------------------------------------------- #
# Set name
layer_name = 'layer_12'
# Set identity
identity = tf.identity(input=net, name=layer_name + '_identity')
# Set layer scope
with tf.variable_scope(layer_name):
# Convolution
net = conv_layer(input_layer=net,
kernel_size=3,
strides=1,
dilation_rate=2,
filters=128,
padding='SAME',
activation=None,
use_bias=False,
name=layer_name + '_conv_ks3_dr128',
seed=self.seed)
# Tanh activation
tanh = tf.nn.tanh(net, name='tanh')
# Sigmoid activation
sigmoid = tf.nn.sigmoid(net, name='sigmoid')
# Combine activations
net = tf.multiply(tanh, sigmoid, name='activation_gate')
# Convolution
net = conv_layer(input_layer=net,
kernel_size=1,
strides=1,
dilation_rate=1,
filters=128,
padding='SAME',
activation=None,
use_bias=False,
name=layer_name + '_conv_ks1_dr1',
seed=self.seed)
# Add identity
net = tf.add(net, identity, name=layer_name + '_identity_add')
# Print shape
print_output_shape(layer_name=layer_name,
net=net,
print_shape=print_shape)
# --- Layer 13 (Convolution) ----------------------------------------------------------------------------- #
# Set name
layer_name = 'layer_13'
# Set identity
identity = tf.identity(input=net, name=layer_name + '_identity')
# Set layer scope
with tf.variable_scope(layer_name):
# Convolution
net = conv_layer(input_layer=net,
kernel_size=3,
strides=1,
dilation_rate=4,
filters=128,
padding='SAME',
activation=None,
use_bias=False,
name=layer_name + '_conv_ks3_dr128',
seed=self.seed)
# Tanh activation
tanh = tf.nn.tanh(net, name='tanh')
# Sigmoid activation
sigmoid = tf.nn.sigmoid(net, name='sigmoid')
# Combine activations
net = tf.multiply(tanh, sigmoid, name='activation_gate')
# Convolution
net = conv_layer(input_layer=net,
kernel_size=1,
strides=1,
dilation_rate=1,
filters=128,
padding='SAME',
activation=None,
use_bias=False,
name=layer_name + '_conv_ks1_dr1',
seed=self.seed)
# Add identity
net = tf.add(net, identity, name=layer_name + '_identity_add')
# Print shape
print_output_shape(layer_name=layer_name,
net=net,
print_shape=print_shape)
# --- Layer 14 (Convolution) ----------------------------------------------------------------------------- #
# Set name
layer_name = 'layer_14'
# Set identity
identity = tf.identity(input=net, name=layer_name + '_identity')
# Set layer scope
with tf.variable_scope(layer_name):
# Convolution
net = conv_layer(input_layer=net,
kernel_size=3,
strides=1,
dilation_rate=8,
filters=128,
padding='SAME',
activation=None,
use_bias=False,
name=layer_name + '_conv_ks3_dr128',
seed=self.seed)
# Tanh activation
tanh = tf.nn.tanh(net, name='tanh')
# Sigmoid activation
sigmoid = tf.nn.sigmoid(net, name='sigmoid')
# Combine activations
net = tf.multiply(tanh, sigmoid, name='activation_gate')
# Convolution
net = conv_layer(input_layer=net,
kernel_size=1,
strides=1,
dilation_rate=1,
filters=128,
padding='SAME',
activation=None,
use_bias=False,
name=layer_name + '_conv_ks1_dr1',
seed=self.seed)
# Add identity
net = tf.add(net, identity, name=layer_name + '_identity_add')
# Print shape
print_output_shape(layer_name=layer_name,
net=net,
print_shape=print_shape)
# --- Layer 15 (Convolution) ----------------------------------------------------------------------------- #
# Set name
layer_name = 'layer_15'
# Set identity
identity = tf.identity(input=net, name=layer_name + '_identity')
# Set layer scope
with tf.variable_scope(layer_name):
# Convolution
net = conv_layer(input_layer=net,
kernel_size=3,
strides=1,
dilation_rate=16,
filters=128,
padding='SAME',
activation=None,
use_bias=False,
name=layer_name + '_conv_ks3_dr128',
seed=self.seed)
# Tanh activation
tanh = tf.nn.tanh(net, name='tanh')
# Sigmoid activation
sigmoid = tf.nn.sigmoid(net, name='sigmoid')
# Combine activations
net = tf.multiply(tanh, sigmoid, name='activation_gate')
# Convolution
net = conv_layer(input_layer=net,
kernel_size=1,
strides=1,
dilation_rate=1,
filters=128,
padding='SAME',
activation=None,
use_bias=False,
name=layer_name + '_conv_ks1_dr1',
seed=self.seed)
# Add identity
net = tf.add(net, identity, name=layer_name + '_identity_add')
# Print shape
print_output_shape(layer_name=layer_name,
net=net,
print_shape=print_shape)
# --- Layer 16 (Convolution) ----------------------------------------------------------------------------- #
# Set name
layer_name = 'layer_16'
# Set identity
identity = tf.identity(input=net, name=layer_name + '_identity')
# Set layer scope
with tf.variable_scope(layer_name):
# Convolution
net = conv_layer(input_layer=net,
kernel_size=3,
strides=1,
dilation_rate=32,
filters=128,
padding='SAME',
activation=None,
use_bias=False,
name=layer_name + '_conv_ks3_dr128',
seed=self.seed)
# Tanh activation
tanh = tf.nn.tanh(net, name='tanh')
# Sigmoid activation
sigmoid = tf.nn.sigmoid(net, name='sigmoid')
# Combine activations
net = tf.multiply(tanh, sigmoid, name='activation_gate')
# Convolution
net = conv_layer(input_layer=net,
kernel_size=1,
strides=1,
dilation_rate=1,
filters=128,
padding='SAME',
activation=None,
use_bias=False,
name=layer_name + '_conv_ks1_dr1',
seed=self.seed)
# Add identity
net = tf.add(net, identity, name=layer_name + '_identity_add')
# Print shape
print_output_shape(layer_name=layer_name,
net=net,
print_shape=print_shape)
"""Network Output"""
# --- Global Average Pooling Layer ----------------------------------------------------------------------- #
# Set name
layer_name = 'gap'
# Set layer scope
with tf.variable_scope(layer_name):
# Reduce mean along dimension 1
gap = tf.reduce_mean(input_tensor=net, axis=1)
# Print shape
print_output_shape(layer_name=layer_name,
net=gap,
print_shape=print_shape)
# --- Softmax Layer -------------------------------------------------------------------------------------- #
# Set name
layer_name = 'logits'
# Softmax activation
logits = fc_layer(input_layer=gap,
neurons=self.classes,
activation=None,
use_bias=False,
name=layer_name,
seed=self.seed)
# Print shape
print_output_shape(layer_name=layer_name,
net=logits,
print_shape=print_shape)
# Compute Class Activation Maps
cams = self._get_cams(net=net, is_training=is_training)
return logits, net, cams
def inference(self,
input_layer,
reuse,
is_training,
name,
print_shape=True):
"""Forward propagation of computational graph."""
# Check input layer dimensions
assert input_layer.shape[1] == self.length
assert input_layer.shape[2] == self.channels
# Define a scope for reusing the variables
with tf.variable_scope(name, reuse=reuse):
"""Block Series 1"""
# --- Layer 1 (Convolution) ------------------------------------------------------------------------------ #
# Set name
layer_name = 'layer_1'
# Set layer scope
with tf.variable_scope(layer_name):
# Convolution
net = conv_layer(input_layer=input_layer,
kernel_size=24,
strides=1,
dilation_rate=1,
filters=128,
padding='SAME',
activation=tf.nn.relu,
use_bias=True,
name=layer_name + '_conv_ks3_dr1',
seed=self.seed)
# Max pool
net = max_pool_layer(input_layer=net,
pool_size=3,
strides=2,
padding='SAME',
name=layer_name + '_maxpool_ps3')
# Print shape
print_output_shape(layer_name=layer_name,
net=net,
print_shape=print_shape)
# --- Layer 2 (Convolution) ------------------------------------------------------------------------------ #
# Set name
layer_name = 'layer_2'
# Set layer scope
with tf.variable_scope(layer_name):
# Convolution
net = conv_layer(input_layer=net,
kernel_size=12,
strides=1,
dilation_rate=2,
filters=128,
padding='SAME',
activation=tf.nn.relu,
use_bias=True,
name=layer_name + '_conv_ks3_dr1',
seed=self.seed)
# Print shape
print_output_shape(layer_name=layer_name,
net=net,
print_shape=print_shape)
# --- Layer 3 (Convolution) ------------------------------------------------------------------------------ #
# Set name
layer_name = 'layer_3'
# Set layer scope
with tf.variable_scope(layer_name):
# Convolution
net = conv_layer(input_layer=net,
kernel_size=12,
strides=1,
dilation_rate=1,
filters=128,
padding='SAME',
activation=tf.nn.relu,
use_bias=True,
name=layer_name + '_conv_ks3_dr1',
seed=self.seed)
# Max pool
net = max_pool_layer(input_layer=net,
pool_size=3,
strides=2,
padding='SAME',
name=layer_name + '_maxpool_ps3')
# Dropout
net = dropout_layer(input_layer=net,
drop_rate=0.3,
seed=self.seed,
training=is_training,
name=layer_name + '_dropout')
# Print shape
print_output_shape(layer_name=layer_name,
net=net,
print_shape=print_shape)
"""Block Series 2"""
# --- Layer 4 (Convolution) ------------------------------------------------------------------------------ #
# Set name
layer_name = 'layer_4'
# Set layer scope
with tf.variable_scope(layer_name):
# Convolution
net = conv_layer(input_layer=net,
kernel_size=6,
strides=1,
dilation_rate=4,
filters=256,
padding='SAME',
activation=tf.nn.relu,
use_bias=True,
name=layer_name + '_conv_ks3_dr1',
seed=self.seed)
# Print shape
print_output_shape(layer_name=layer_name,
net=net,
print_shape=print_shape)
# --- Layer 5 (Convolution) ------------------------------------------------------------------------------ #
# Set name
layer_name = 'layer_5'
# Set layer scope
with tf.variable_scope(layer_name):
# Convolution
net = conv_layer(input_layer=net,
kernel_size=6,
strides=1,
dilation_rate=1,
filters=256,
padding='SAME',
activation=tf.nn.relu,
use_bias=True,
name=layer_name + '_conv_ks3_dr1',
seed=self.seed)
# Print shape
print_output_shape(layer_name=layer_name,
net=net,
print_shape=print_shape)
# --- Layer 6 (Convolution) ------------------------------------------------------------------------------ #
# Set name
layer_name = 'layer_6'
# Set layer scope
with tf.variable_scope(layer_name):
# Convolution
net = conv_layer(input_layer=net,
kernel_size=6,
strides=1,
dilation_rate=8,
filters=256,
padding='SAME',
activation=tf.nn.relu,
use_bias=True,
name=layer_name + '_conv_ks3_dr1',
seed=self.seed)
# Max pool
net = max_pool_layer(input_layer=net,
pool_size=3,
strides=2,
padding='SAME',
name=layer_name + '_maxpool_ps3')
# Dropout
net = dropout_layer(input_layer=net,
drop_rate=0.3,
seed=self.seed,
training=is_training,
name=layer_name + '_dropout')
# Print shape
print_output_shape(layer_name=layer_name,
net=net,
print_shape=print_shape)
"""Block Series 3"""
# --- Layer 7 (Convolution) ------------------------------------------------------------------------------ #
# Set name
layer_name = 'layer_7'
# Set layer scope
with tf.variable_scope(layer_name):
# Convolution
net = conv_layer(input_layer=net,
kernel_size=3,
strides=1,
dilation_rate=1,
filters=512,
padding='SAME',
activation=tf.nn.relu,
use_bias=True,
name=layer_name + '_conv_ks3_dr1',
seed=self.seed)
# Print shape
print_output_shape(layer_name=layer_name,
net=net,
print_shape=print_shape)
# --- Layer 8 (Convolution) ------------------------------------------------------------------------------ #
# Set name
layer_name = 'layer_8'
# Set layer scope
with tf.variable_scope(layer_name):
# Convolution
net = conv_layer(input_layer=net,
kernel_size=3,
strides=1,
dilation_rate=16,
filters=512,
padding='SAME',
activation=tf.nn.relu,
use_bias=True,
name=layer_name + '_conv_ks3_dr1',
seed=self.seed)
# Print shape
print_output_shape(layer_name=layer_name,
net=net,
print_shape=print_shape)
# --- Layer 9 (Convolution) ------------------------------------------------------------------------------ #
# Set name
layer_name = 'layer_9'
# Set layer scope
with tf.variable_scope(layer_name):
# Convolution
net = conv_layer(input_layer=net,
kernel_size=3,
strides=1,
dilation_rate=1,
filters=512,
padding='SAME',
activation=tf.nn.relu,
use_bias=True,
name=layer_name + '_conv_ks3_dr1',
seed=self.seed)
# Dropout
net = dropout_layer(input_layer=net,
drop_rate=0.3,
seed=self.seed,
training=is_training,
name=layer_name + '_dropout')
# Print shape
print_output_shape(layer_name=layer_name,
net=net,
print_shape=print_shape)
"""Block Series 4"""
# --- Layer 10 (Convolution) ----------------------------------------------------------------------------- #
# Set name
layer_name = 'layer_10'
# Set layer scope
with tf.variable_scope(layer_name):
# Convolution
net = conv_layer(input_layer=net,
kernel_size=3,
strides=1,
dilation_rate=32,
filters=512,
padding='SAME',
activation=tf.nn.relu,
use_bias=True,
name=layer_name + '_conv_ks3_dr1',
seed=self.seed)
# Print shape
print_output_shape(layer_name=layer_name,
net=net,
print_shape=print_shape)
# --- Layer 11 (Convolution) ----------------------------------------------------------------------------- #
# Set name
layer_name = 'layer_11'
# Set layer scope
with tf.variable_scope(layer_name):
# Convolution
net = conv_layer(input_layer=net,
kernel_size=3,
strides=1,
dilation_rate=1,
filters=512,
padding='SAME',
activation=tf.nn.relu,
use_bias=True,
name=layer_name + '_conv_ks3_dr1',
seed=self.seed)
# Print shape
print_output_shape(layer_name=layer_name,
net=net,
print_shape=print_shape)
# --- Layer 12 (Convolution) ----------------------------------------------------------------------------- #
# Set name
layer_name = 'layer_12'
# Set layer scope
with tf.variable_scope(layer_name):
# Convolution
net = conv_layer(input_layer=net,
kernel_size=3,
strides=1,
dilation_rate=64,
filters=512,
padding='SAME',
activation=tf.nn.relu,
use_bias=True,
name=layer_name + '_conv_ks3_dr1',
seed=self.seed)
# Dropout
net = dropout_layer(input_layer=net,
drop_rate=0.3,
seed=self.seed,
training=is_training,
name=layer_name + '_dropout')
# Print shape
print_output_shape(layer_name=layer_name,
net=net,
print_shape=print_shape)
"""Network Output"""
# --- Global Average Pooling Layer ----------------------------------------------------------------------- #
# Set name
layer_name = 'gap'
# Set layer scope
with tf.variable_scope(layer_name):
# Reduce mean along dimension 1
gap = tf.reduce_mean(input_tensor=net, axis=1)
# Print shape
print_output_shape(layer_name=layer_name,
net=gap,
print_shape=print_shape)
# --- Softmax Layer -------------------------------------------------------------------------------------- #
# Set name
layer_name = 'logits'
# Softmax activation
logits = fc_layer(input_layer=gap,
neurons=self.classes,
activation=None,
use_bias=False,
name=layer_name,
seed=self.seed)
# Print shape
print_output_shape(layer_name=layer_name,
net=logits,
print_shape=print_shape)
# Compute Class Activation Maps
cams = self._get_cams(net=net, is_training=is_training)
return logits, net, cams