def build_top_layer_model(base_layer,num_depthwise_layer=None,
num_fully_connected_layer=1,num_hidden_unit=512,
activation_fully_connected='relu',dropout_keep_prob=None,regularizers=None,num_classes=1000):
previous_layer = base_layer
if num_depthwise_layer!=None:
num_depthwise_layer = num_depthwise_layer * 3
for i in range(num_depthwise_layer):
depth_wise_net = depthwise_convolution_2d(input_tensor=previous_layer,number_filters=1,
kernel_sizes=(3,3), stride_sizes=(2,2), paddings='same',
activations='relu',names=str(i))
previous_layer = depth_wise_net
flatten_layer = flatten(input_tensor=previous_layer)
if num_fully_connected_layer !=None:
for j in range(num_fully_connected_layer):
fully_connected_net = dense(input_tensor=flatten_layer,hidden_units=num_hidden_unit,
activations=activation_fully_connected,regularizers=regularizers,
scale=dropout_keep_prob,names=str(j))
flatten_layer = fully_connected_net
else:
flatten_layer = flatten_layer
non_logit = dense(input_tensor=flatten_layer,hidden_units=num_classes,activations=None)
if num_classes > 2:
output = softmax(input_tensor=non_logit, names='output')
else:
output = sigmoid(input_tensor=non_logit, names='output')
return non_logit, output
def create_model(self):
number_filter = self.input_tensor.get_shape().as_list()[-1]
vgg_base = self.vgg_block(input_tensor=self.input_tensor, num_block=self.num_block, batch_normalization=self.batch_normalization)
base_var_list = tf.compat.v1.get_collection(tf.compat.v1.GraphKeys.GLOBAL_VARIABLES)
with tf.compat.v1.variable_scope(self.scope, 'vgg_16', [vgg_base]):
if self.num_depthwise_layer!=None or self.num_depthwise_layer>0:
for j in range(0,self.num_depthwise_layer):
##### FIX THIS TOMORROW
vgg_base = depthwise_convolution_2d(input_tensor=vgg_base, number_filters=number_filter, kernel_sizes=(3,3), stride_sizes=(1,1), paddings='same', activations='relu', dropout_rates=None, names=None)
else:
flatten_layer = flatten(input_tensor=vgg_base, names='flatten')
for i in range(0, self.num_dense_layer):
vgg_base = dense(input_tensor=flatten_layer, hidden_units=self.num_hidden_unit, activations=self.activation_dense, regularizers=self.regularizer, scale=self.dropout_rate)
last_layer = flatten(input_tensor=vgg_base, names='flatten')
full_var_list = tf.compat.v1.get_collection(tf.compat.v1.GraphKeys.GLOBAL_VARIABLES)
non_logit = dense(input_tensor=last_layer, hidden_units=self.classes, names='output')
if self.classes > 2:
output = softmax(input_tensor=non_logit, names='output')
else:
output = sigmoid(input_tensor=non_logit, names='output')
return non_logit, output, base_var_list, full_var_list
def create_model(self):
"""[summary]
Arguments:
input_tensor {[type]} -- [description]
Keyword Arguments:
classes {int} -- [description] (default: {1000})
n_inception_a {int} -- [description] (default: {4})
n_inception_b {int} -- [description] (default: {7})
n_inception_c {int} -- [description] (default: {3})
batch_normalizations {bool} -- [description] (default: {True})
"""
with tf.compat.v1.variable_scope(self.scope, 'inception_v4_resnet_v2', [self.input_tensor]):
stem_layer = self.stem_block(input_tensor=self.input_tensor, batch_normalization=self.batch_normalizations)
for i in range(0,self.n_inception_a):
inception_a_layer = self.inception_a(input_tensor=stem_layer, batch_normalization=self.batch_normalizations)
reduction_a_layer = self.reduction_a(input_tensor=inception_a_layer, batch_normalization=self.batch_normalizations)
for j in range(0,self.n_inception_b):
inception_b_layer = self.inception_b(input_tensor=reduction_a_layer, batch_normalization=self.batch_normalizations)
reduction_b_layer = self.reduction_b(input_tensor=inception_b_layer, batch_normalization=self.batch_normalizations)
for k in range(0,self.n_inception_c):
inception_c_layer = self.inception_c(input_tensor=reduction_b_layer, batch_normalization=self.batch_normalizations)
if n_inception_a == 0:
inception_v4 = stem_layer
elif n_inception_b == 0:
inception_v4 = reduction_a_layer
elif n_inception_c == 0:
inception_v4 = reduction_b_layer
else:
inception_v4 = inception_c_layer
base_var_list = tf.compat.v1.get_collection(tf.compat.v1.GraphKeys.GLOBAL_VARIABLES)
inception_v4 = avarage_pool_2d(input_tensor=inception_v4, kernel_sizes=(3,3), paddings='SAME', stride_sizes=(1,1), names='output')
inception_v4 = dropout(input_tensor=inception_v4, names='output', dropout_rates=0.2)
inception_v4 = flatten(input_tensor=inception_v4, names='output')
full_var_list = tf.compat.v1.get_collection(tf.compat.v1.GraphKeys.GLOBAL_VARIABLES)
non_logit = dense(input_tensor=inception_v4, hidden_units=self.classes, names='output')
if self.classes > 2:
output = softmax(input_tensor=non_logit, names='output')
else:
output = sigmoid(input_tensor=non_logit, names='output')
return non_logit, output, base_var_list, full_var_list
def build_top_layer_model(base_layer,
num_depthwise_layer=None,
dropout_keep_prob=None,
regularizers=None,
num_classes=1000):
previous_layer = base_layer
if num_depthwise_layer != None:
num_depthwise_layer = num_depthwise_layer
for i in range(num_depthwise_layer):
depth_wise_net = depthwise_convolution_2d(
input_tensor=previous_layer,
number_filters=1,
kernel_sizes=(3, 3),
stride_sizes=(2, 2),
paddings='same',
activations='relu',
names=str(i))
previous_layer = depth_wise_net
else:
depth_wise_net = previous_layer
flatten_layer = dropout(input_tensor=flatten_layer,
names='output',
dropout_rates=dropout_keep_prob)
flatten_layer = flatten(input_tensor=depth_wise_net)
full_var_list = tf.compat.v1.get_collection(
tf.compat.v1.GraphKeys.GLOBAL_VARIABLES)
non_logit = dense(input_tensor=flatten_layer,
hidden_units=num_classes,
activations=None)
if num_classes > 2:
output = softmax(input_tensor=non_logit, names='output')
else:
output = sigmoid(input_tensor=non_logit, names='output')
return non_logit, output, full_var_list