def build_discriminator(self):
d_input = Input(shape=(self.latent_dim,))
layers = []
layers.append(Dense(units=self.latent_dim, input_dim=self.latent_dim))
layers.append(LeakyReLU(alpha=0.2))
layers.append(Dense(self.latent_dim//2))
layers.append(LeakyReLU(alpha=0.2))
layers.append(Dense(1, activation='sigmoid'))
return d_input, layers
def basic_cnn(word_size=300,
sentence_size=100,
num_filters=100,
filter_sizes=[2, 4, 6]):
model = Sequential()
layers = []
for f in range(len(filter_sizes)):
if f == 0:
layers.append(
Conv1D(num_filters,
kernel_size=filter_sizes[f],
padding='valid',
activation='relu',
kernel_initializer='he_uniform',
input_shape=(sentence_size, word_size)))
else:
layers.append(
Conv1D(num_filters,
kernel_size=filter_sizes[f],
padding='valid',
activation='relu',
kernel_initializer='he_uniform'))
layers.append(MaxPooling1D(pool_size=2))
layers.append(Dropout(0.5))
layers.append(Flatten())
layers.append(Dense(100, activation='relu'))
layers.append(Dropout(0.5))
# layers.append(Dense(50, activation='relu'))
# layers.append(Dropout(0.5))
layers.append(Dense(2, activation='softmax'))
for layer in layers:
model.add(layer)
return model
def build_generator(self):
noise = Input(shape=(self.noise_dims,))
output_dim = self.latent_dim
layers = []
a = LeakyReLU(alpha=0.2)
# a = Activation('relu')
layers.append(Dense(self.noise_dims))
layers.append(a)
layers.append(BatchNormalization(momentum=0.8))
layers.append(Dense(output_dim))
layers.append(a)
layers.append(BatchNormalization(momentum=0.8))
layers.append(Dense(output_dim))
layers.append(a)
layers.append(BatchNormalization(momentum=0.8))
layers.append(Dense(output_dim, activation='tanh'))
return noise, layers
def create_shared_inceptionV4_network_2():
kwargs = {
'padding': 'same',
'dropout': 0.0,
'BN': True,
'kernel_initializer': 'glorot_uniform'
}
input = []
input.append(Input(shape=(350, 38, 1), name='Wire_1'))
input.append(Input(shape=(350, 38, 1), name='Wire_2'))
layers = []
layers.append(Conv_block(32, filter_size=(3, 3), name='1', **kwargs))
layers.append(Conv_block(32, filter_size=(3, 3), name='2', **kwargs))
layers.append([MaxPooling2D((4, 1), name='3')]) #(4,2)
layers.append(Conv_block(64, filter_size=(3, 3), name='4', **kwargs))
num_filters = (96, (64, 96), (64, 96), 96
) #TODO Real values from IncV1 Paper?
layers.append(InceptionV4_block(num_filters=num_filters, name='5'))
layers.append(InceptionV4_block(num_filters=num_filters, name='6'))
layers.append([MaxPooling2D((2, 2), name='7')])
layers.append(InceptionV4_block(num_filters=num_filters, name='8'))
layers.append(InceptionV4_block(num_filters=num_filters, name='9'))
layers.append([MaxPooling2D((2, 1), name='10')])
layers.append(InceptionV4_block(num_filters=num_filters, name='11'))
layers.append(InceptionV4_block(num_filters=num_filters, name='12'))
layers.append([GlobalAveragePooling2D(name='18')])
paths = assemble_network(input, layers)
merge = Concatenate(name='Conc_Top_1_and_2')(paths)
output = Dense(2,
name='Output_Top',
activation='softmax',
kernel_initializer="glorot_uniform")(merge)
return Model(inputs=input, outputs=output)
def create_shared_inception_network_4():
kwargs = {
'padding': 'same',
'dropout': 0.0,
'BN': True,
'kernel_initializer': 'glorot_uniform'
}
inputU = []
inputU.append(Input(shape=(350, 38, 1), name='U-Wire_1'))
inputU.append(Input(shape=(350, 38, 1), name='U-Wire_2'))
inputV = []
inputV.append(Input(shape=(350, 38, 1), name='V-Wire_1'))
inputV.append(Input(shape=(350, 38, 1), name='V-Wire_2'))
layersU = []
layersV = []
for layers in [
layersU, layersV
]: # Use same architecture for U and V wires. Can be changed
layers.append(Conv_block(32, filter_size=(3, 3), **kwargs))
layers.append(Conv_block(32, filter_size=(3, 3), **kwargs))
layers.append([MaxPooling2D((4, 2))])
layers.append(Conv_block(64, filter_size=(3, 3), **kwargs))
num_filters = (64, (96, 128), (16, 32), 32
) #TODO Real values from IncV1 Paper?
num_filters_deep = (96, (96, 192), (64, 96), 96)
layers.append(InceptionV1_block(num_filters=num_filters))
layers.append(InceptionV1_block(num_filters=num_filters))
layers.append([
MaxPooling2D((2, 2))
]) # TODO Test Inception module with stride=2 instead of max pooling
layers.append(InceptionV1_block(num_filters=num_filters))
layers.append(InceptionV1_block(num_filters=num_filters))
layers.append([MaxPooling2D((2, 1))])
layers.append(InceptionV1_block(num_filters=num_filters))
layers.append(InceptionV1_block(num_filters=num_filters))
# layers.append([MaxPooling2D((2, 1))])
# layers.append(InceptionV1_block(num_filters=num_filters))
# layers.append(InceptionV1_block(num_filters=num_filters))
# layers.append(InceptionV1_block(num_filters=num_filters_deep))
# layers.append(InceptionV1_block(num_filters=num_filters_deep))
layers.append([GlobalAveragePooling2D()])
pathsU = assemble_network(inputU, layersU)
pathsV = assemble_network(inputV, layersV)
inputUV = []
inputUV.append(Concatenate(name='TPC_1')([pathsU[0], pathsV[0]]))
inputUV.append(Concatenate(name='TPC_2')([pathsU[1], pathsV[1]]))
layersUV = []
layersUV.append([Dense(64, activation='relu')])
layersUV.append([Dense(16, activation='relu')])
pathsUV = assemble_network(inputUV, layersUV)
merge = Concatenate(name='Flat_1_and_2')(pathsUV)
output = Dense(2,
name='Output',
activation='softmax',
kernel_initializer="glorot_uniform")(merge)
inputUV = []
inputUV.append(inputU[0])
inputUV.append(inputV[0])
inputUV.append(inputU[1])
inputUV.append(inputV[1])
return Model(inputs=inputUV, outputs=output)
def create_shared_inception_network_2():
# kwargs = {'padding': 'same',
# 'dropout': 0.2, #TODO TEST DROPOUT
# 'BN': True,
# 'kernel_initializer': 'glorot_uniform'}
#
# input = []
# input.append(Input(shape=(350, 38, 1), name='Wire_1'))
# input.append(Input(shape=(350, 38, 1), name='Wire_2'))
#
# layers = []
# layers.append(Conv_block(32, filter_size=(3, 3), name='1', **kwargs))
# layers.append(Conv_block(32, filter_size=(3, 3), name='2', **kwargs))
# layers.append([MaxPooling2D((4, 2), name='3')])
# layers.append(Conv_block(64, filter_size=(3, 3), name='4', **kwargs))
#
# num_filters = (64, (96, 128), (16, 32), 32) #TODO Real values from IncV1 Paper?
# layers.append(InceptionV1_block(num_filters=num_filters, name='5'))
# layers.append(InceptionV1_block(num_filters=num_filters, name='6'))
# layers.append([MaxPooling2D((2, 2), name='7')]) # TODO Test Inception module with stride=2 instead of max pooling
# layers.append(InceptionV1_block(num_filters=num_filters, name='8'))
# layers.append(InceptionV1_block(num_filters=num_filters, name='9'))
# layers.append([MaxPooling2D((2, 1), name='10')])
# layers.append(InceptionV1_block(num_filters=num_filters, name='11'))
# layers.append(InceptionV1_block(num_filters=num_filters, name='12'))
# layers.append([MaxPooling2D((2, 1), name='13')])
# layers.append(InceptionV1_block(num_filters=num_filters, name='14'))
# layers.append(InceptionV1_block(num_filters=num_filters, name='15'))
# layers.append(InceptionV1_block(num_filters=num_filters, name='16'))
# layers.append(InceptionV1_block(num_filters=num_filters, name='17'))
#
# layers.append([GlobalAveragePooling2D(name='18')])
#
# paths = assemble_network(input, layers)
#
# merge = Concatenate(name='Conc_Top_1_and_2')(paths)
# output = Dense(2, name='Output_Top', activation='softmax', kernel_initializer="glorot_uniform")(merge)
#
# return Model(inputs=input, outputs=output)
# TODO Baseline below
kwargs = {
'padding': 'same',
'dropout': 0.0,
'BN': True,
'kernel_initializer': 'glorot_uniform'
}
input = []
input.append(Input(shape=(350, 38, 1), name='Wire_1'))
input.append(Input(shape=(350, 38, 1), name='Wire_2'))
layers = []
layers.append(Conv_block(32, filter_size=(3, 3), name='1', **kwargs))
layers.append(Conv_block(32, filter_size=(3, 3), name='2', **kwargs))
layers.append([MaxPooling2D((4, 2), name='3')])
layers.append(Conv_block(64, filter_size=(3, 3), name='4', **kwargs))
num_filters = (64, (96, 128), (16, 32), 32
) #TODO Real values from IncV1 Paper?
layers.append(InceptionV1_block(num_filters=num_filters, name='5'))
layers.append(InceptionV1_block(num_filters=num_filters, name='6'))
layers.append([
MaxPooling2D((2, 2), name='7')
]) # TODO Test Inception module with stride=2 instead of max pooling
layers.append(InceptionV1_block(num_filters=num_filters, name='8'))
layers.append(InceptionV1_block(num_filters=num_filters, name='9'))
layers.append([MaxPooling2D((2, 1), name='10')])
layers.append(InceptionV1_block(num_filters=num_filters, name='11'))
layers.append(InceptionV1_block(num_filters=num_filters, name='12'))
layers.append([MaxPooling2D((2, 1), name='13')])
layers.append(InceptionV1_block(num_filters=num_filters, name='14'))
layers.append(InceptionV1_block(num_filters=num_filters, name='15'))
layers.append(InceptionV1_block(num_filters=num_filters, name='16'))
layers.append(InceptionV1_block(num_filters=num_filters, name='17'))
layers.append([GlobalAveragePooling2D(name='18')])
paths = assemble_network(input, layers)
merge = Concatenate(name='Conc_Top_1_and_2')(paths)
output = Dense(2,
name='Output_Top',
activation='softmax',
kernel_initializer="glorot_uniform")(merge)
return Model(inputs=input, outputs=output)
def create_shared_inception_network_2_extra_input(kwargs_inc={}):
kwargs_out = dict(kwargs_inc)
for key in ['dropout', 'activity_regularizer', 'kernel_regularizer']:
kwargs_out.pop(key, None)
# auxiliary_input = Input(shape=(10, 4), name='Aux_Input')
# x = LocallyConnected1D(10, kernel_size=1, activation='relu', name='21', **kwargs_inc)(auxiliary_input)
# x = LocallyConnected1D(20, kernel_size=1, activation='relu', name='22', **kwargs_inc)(x)
# # x = LocallyConnected1D(40, kernel_size=1, activation='relu', name='21')(x)
# merge_pos = Flatten(name='Flatten_Pos')(x)
# output_pos = Dense(2, name='Output_Pos', activation='softmax', kernel_initializer="glorot_uniform", **kwargs_out)(merge_pos)
# # return Model(inputs=[auxiliary_input], outputs=[output_pos])
# auxiliary_input = Input(shape=(10, 4, 1), name='aux_input')
# x = Conv2D(10, kernel_size=(1, 4), padding='same', activation='relu', kernel_initializer="glorot_normal")(auxiliary_input)
# # # x = Conv2D(10, kernel_size=(1, 4), padding='same', activation='relu', kernel_initializer="glorot_uniform")(x)
# # # x = Conv2D(10, kernel_size=(1, 4), padding='same', activation='relu', kernel_initializer="glorot_uniform")(x)
# # # x = Conv2D(20, kernel_size=(1, 4), padding='same', activation='relu', kernel_initializer="glorot_uniform")(x)
# x = Conv2D(20, kernel_size=(1, 4), padding='same', activation='relu', kernel_initializer="glorot_normal")(x)
# x = Conv2D(40, kernel_size=(1, 4), padding='same', activation='relu', kernel_initializer="glorot_normal")(x)
# merge_pos = Flatten()(x)
# # # merge_pos = Dense(56, activation='relu', kernel_initializer="glorot_uniform")(x)
# output_pos = Dense(2, name='Output_Pos', activation='softmax', kernel_initializer="glorot_normal")(merge_pos)
# return Model(inputs=[auxiliary_input], outputs=[output_pos])
kwargs = {
'padding': 'same',
'dropout': 0.0,
'BN': True,
'kernel_initializer': 'glorot_uniform'
}
kwargs = merge_two_dicts(kwargs, kwargs_inc)
input = []
input.append(Input(shape=(350, 38, 1), name='Wire_1'))
input.append(Input(shape=(350, 38, 1), name='Wire_2'))
layers = []
layers.append(Conv_block(32, filter_size=(3, 3), name='ConvBl_1',
**kwargs))
layers.append(Conv_block(32, filter_size=(3, 3), name='ConvBl_2',
**kwargs))
layers.append([MaxPooling2D((4, 2), name='maxp_1')])
layers.append(Conv_block(64, filter_size=(3, 3), name='ConvBl_3',
**kwargs))
num_filters = (64, (96, 128), (16, 32), 32
) #TODO Real values from IncV1 Paper?
layers.append(
InceptionV1_block(num_filters=num_filters,
name='IncBl_1',
kwargs_inc=kwargs_inc))
layers.append(
InceptionV1_block(num_filters=num_filters,
name='IncBl_2',
kwargs_inc=kwargs_inc))
layers.append([
MaxPooling2D((2, 2), name='maxp_2')
]) # TODO Test Inception module with stride=2 instead of max pooling
layers.append(
InceptionV1_block(num_filters=num_filters,
name='IncBl_3',
kwargs_inc=kwargs_inc))
layers.append(
InceptionV1_block(num_filters=num_filters,
name='IncBl_4',
kwargs_inc=kwargs_inc))
layers.append([MaxPooling2D((2, 1), name='maxp_3')])
layers.append(
InceptionV1_block(num_filters=num_filters,
name='IncBl_5',
kwargs_inc=kwargs_inc))
layers.append(
InceptionV1_block(num_filters=num_filters,
name='IncBl_6',
kwargs_inc=kwargs_inc))
layers.append([MaxPooling2D((2, 1), name='maxp_4')])
layers.append(
InceptionV1_block(num_filters=num_filters,
name='IncBl_7',
kwargs_inc=kwargs_inc))
layers.append(
InceptionV1_block(num_filters=num_filters,
name='IncBl_8',
kwargs_inc=kwargs_inc))
layers.append(
InceptionV1_block(num_filters=num_filters,
name='IncBl_9',
kwargs_inc=kwargs_inc))
layers.append(
InceptionV1_block(num_filters=num_filters,
name='IncBl_10',
kwargs_inc=kwargs_inc))
layers.append([GlobalAveragePooling2D(name='GAverPool_Top')])
paths = assemble_network(input, layers)
merge_top = Concatenate(name='Concat_Top_1_and_2')(paths)
output_top = Dense(2,
name='Output_Top',
activation='softmax',
kernel_initializer="glorot_uniform",
**kwargs_out)(merge_top)
# return Model(inputs=input, outputs=output_top)
auxiliary_input = Input(shape=(10, 4), name='Aux_Input')
auxiliary_flat = Flatten(name='Flatten_Aux_Input')(auxiliary_input)
x = Concatenate(name='Merge_Pos_and_Top')([merge_top, auxiliary_flat])
# x = Dense(64, activation='relu', kernel_initializer="glorot_uniform", name='31')(x)
# x = Dense(64, activation='relu', kernel_initializer="glorot_uniform", name='32')(x)
# x = Dense(64, activation='relu', kernel_initializer="glorot_uniform", name='33')(x)
x = Dense(256,
activation='relu',
kernel_initializer="glorot_uniform",
name='31')(x)
x = Dropout(0.3, name='drop_1')(x)
x = Dense(64,
activation='relu',
kernel_initializer="glorot_uniform",
name='32')(x)
x = Dropout(0.3, name='drop_2')(x)
x = Dense(16,
activation='relu',
kernel_initializer="glorot_uniform",
name='33')(x)
output = Dense(2,
name='Output',
activation='softmax',
kernel_initializer="glorot_uniform")(x)
return Model(inputs=[input[0], input[1], auxiliary_input],
outputs=[output, output_top])
def create_shared_dcnn_network_2():
kwargs = {
'padding': 'same',
'dropout': 0.0,
'BN': True,
'kernel_initializer': 'glorot_uniform'
}
# 'kernel_regularizer': regularizers.l2(1.e-2)}
input = []
input.append(Input(shape=(350, 38, 1), name='Wire_1'))
input.append(Input(shape=(350, 38, 1), name='Wire_2'))
layers = []
layers.append(
Conv_block(16,
filter_size=(5, 3),
name='ConvBl_1',
max_pooling=None,
**kwargs))
layers.append(
Conv_block(16,
filter_size=(5, 3),
name='ConvBl_2',
max_pooling=(4, 2),
**kwargs))
layers.append(
Conv_block(32,
filter_size=(5, 3),
name='ConvBl_3',
max_pooling=None,
**kwargs))
layers.append(
Conv_block(32,
filter_size=(5, 3),
name='ConvBl_4',
max_pooling=(4, 2),
**kwargs))
layers.append(
Conv_block(64,
filter_size=(3, 3),
name='ConvBl_5',
max_pooling=None,
**kwargs))
layers.append(
Conv_block(64,
filter_size=(3, 3),
name='ConvBl_6',
max_pooling=(2, 2),
**kwargs))
layers.append(
Conv_block(128,
filter_size=(3, 3),
name='ConvBl_7',
max_pooling=None,
**kwargs))
layers.append(
Conv_block(128,
filter_size=(3, 3),
name='ConvBl_8',
max_pooling=(2, 2),
**kwargs))
layers.append(
Conv_block(256,
filter_size=(3, 3),
name='ConvBl_9',
max_pooling=None,
**kwargs))
#TEST
layers.append(
Conv_block(256,
filter_size=(3, 3),
name='ConvBl_10',
max_pooling=None,
**kwargs))
layers.append([GlobalAveragePooling2D()])
#TEST
# layers.append(Conv_block(256, k_size=(3, 3), padding=padding, init=init, dropout=drop, max_pooling=(2, 2), activ=act, kernel_reg=regu))
# layers.append([Flatten()])
# layers.append([Dense(64, activation='relu', kernel_initializer=kwargs['kernel_initializer'])]) #, kernel_regularizer=regu)])
# layers.append([Dense(32, activation='relu', kernel_initializer=kwargs['kernel_initializer'])]) #, kernel_regularizer=regu)])
paths = assemble_network(input, layers)
merge = Concatenate(name='Flat_1_and_2')(paths)
# merge = paths[0]
output = Dense(2,
name='Output',
activation='softmax',
kernel_initializer=kwargs['kernel_initializer'])(merge)
return Model(inputs=input, outputs=output)
def create_shared_dcnn_network_4():
kwargs = {
'padding': 'same',
'dropout': 0.0,
'BN': False,
'kernel_initializer': 'glorot_uniform',
'kernel_regularizer': regularizers.l2(1.e-2)
}
inputU = []
inputU.append(Input(shape=(350, 38, 1), name='U-Wire_1'))
inputU.append(Input(shape=(350, 38, 1), name='U-Wire_2'))
inputV = []
inputV.append(Input(shape=(350, 38, 1), name='V-Wire_1'))
inputV.append(Input(shape=(350, 38, 1), name='V-Wire_2'))
layersU = []
layersV = []
for layers in [
layersU, layersV
]: # Use same architecture for U and V wires. Can be changed
layers.append(
Conv_block(16, filter_size=(5, 3), max_pooling=None, **kwargs))
layers.append(
Conv_block(16, filter_size=(5, 3), max_pooling=(4, 2), **kwargs))
layers.append(
Conv_block(32, filter_size=(5, 3), max_pooling=None, **kwargs))
layers.append(
Conv_block(32, filter_size=(5, 3), max_pooling=(4, 2), **kwargs))
layers.append(
Conv_block(64, filter_size=(3, 3), max_pooling=None, **kwargs))
layers.append(
Conv_block(64, filter_size=(3, 3), max_pooling=(2, 2), **kwargs))
layers.append(
Conv_block(128, filter_size=(3, 3), max_pooling=None, **kwargs))
layers.append(
Conv_block(128, filter_size=(3, 3), max_pooling=(2, 2), **kwargs))
layers.append(
Conv_block(256, filter_size=(3, 3), max_pooling=None, **kwargs))
layers.append(
Conv_block(256, filter_size=(3, 3), max_pooling=None, **kwargs))
layers.append([GlobalAveragePooling2D()])
pathsU = assemble_network(inputU, layersU)
pathsV = assemble_network(inputV, layersV)
inputUV = []
inputUV.append(Concatenate(name='TPC_1')([pathsU[0], pathsV[0]]))
inputUV.append(Concatenate(name='TPC_2')([pathsU[1], pathsV[1]]))
layersUV = []
layersUV.append([
Dense(32,
activation='relu',
kernel_regularizer=kwargs['kernel_regularizer'])
])
layersUV.append([
Dense(16,
activation='relu',
kernel_regularizer=kwargs['kernel_regularizer'])
])
pathsUV = assemble_network(inputUV, layersUV)
merge = Concatenate(name='Flat_1_and_2')(pathsUV)
output = Dense(2,
name='Output',
activation='softmax',
kernel_initializer=kwargs['kernel_initializer'])(merge)
inputUV = []
inputUV.append(inputU[0])
inputUV.append(inputV[0])
inputUV.append(inputU[1])
inputUV.append(inputV[1])
return Model(inputs=inputUV, outputs=output)
def build_generator(self):
noise = Input(shape=(self.noise_dims, ))
output_dim = self.latent_dim
layers = []
a = LeakyReLU(alpha=0.2)
# a = Activation('relu')
layers.append(Dense(self.noise_dims))
layers.append(a)
layers.append(BatchNormalization(momentum=0.8))
layers.append(Dense(output_dim))
layers.append(a)
layers.append(BatchNormalization(momentum=0.8))
layers.append(Dense(output_dim))
layers.append(a)
layers.append(BatchNormalization(momentum=0.8))
layers.append(Dense(output_dim, activation='tanh'))
return noise, layers
