data_format='channels_last')(_inputBN)
C1 = LeakyReLU()(C1)
C2 = ConvSN2D(25, (7, 7), padding='valid', data_format='channels_last')(C1)
C2 = LeakyReLU()(C2)
C3 = ConvSN2D(40, (9, 9), padding='valid', data_format='channels_last')(C2)
C3 = LeakyReLU()(C3)
C4 = ConvSN2D(50, (11, 11), padding='valid', data_format='channels_last')(C3)
C4 = LeakyReLU()(C4)
Average_score = GlobalAveragePooling2D(name='Average_score')(
C4) #(batch_size, channels)
D1 = DenseSN(50)(Average_score)
D1 = LeakyReLU()(D1)
D2 = DenseSN(10)(D1)
D2 = LeakyReLU()(D2)
Score = DenseSN(1)(D2)
Discriminator = Model(outputs=Score, inputs=_input)
Discriminator.trainable = True
Discriminator.compile(loss='mse', optimizer='adam')
#### Combine the two networks to become MetricGAN
Discriminator.trainable = False
def BuildDiscriminator(summary=True,
spectral_normalization=True,
batch_normalization=False,
bn_momentum=0.9,
bn_epsilon=0.00002,
resnet=True,
name='Discriminator',
plot=False):
if resnet:
model_input = Input(shape=(32, 32, 3))
resblock_1 = ResBlock(input_shape=(32, 32, 3),
channels=128,
sampling='down',
batch_normalization=True,
spectral_normalization=spectral_normalization,
name='Discriminator_resblock_Down_1')
h = resblock_1(model_input)
resblock_2 = ResBlock(input_shape=(16, 16, 128),
channels=128,
sampling='down',
batch_normalization=True,
spectral_normalization=spectral_normalization,
name='Discriminator_resblock_Down_2')
h = resblock_2(h)
resblock_3 = ResBlock(input_shape=(8, 8, 128),
channels=128,
sampling=None,
batch_normalization=True,
spectral_normalization=spectral_normalization,
trainable_sortcut=False,
name='Discriminator_resblock_1')
h = resblock_3(h)
resblock_4 = ResBlock(input_shape=(8, 8, 128),
channels=128,
sampling=None,
batch_normalization=True,
spectral_normalization=spectral_normalization,
trainable_sortcut=False,
name='Discriminator_resblock_2')
h = resblock_4(h)
h = Activation('relu')(h)
h = GlobalSumPooling2D()(h)
model_output = DenseSN(1, kernel_initializer='glorot_uniform')(h)
model = Model(model_input, model_output, name=name)
else:
if spectral_normalization:
model = Sequential(name=name)
model.add(
ConvSN2D(64,
kernel_size=3,
strides=1,
kernel_initializer='glorot_uniform',
padding='same',
input_shape=(32, 32, 3)))
model.add(LeakyReLU(0.1))
model.add(
ConvSN2D(64,
kernel_size=4,
strides=2,
kernel_initializer='glorot_uniform',
padding='same'))
model.add(LeakyReLU(0.1))
model.add(
ConvSN2D(128,
kernel_size=3,
strides=1,
kernel_initializer='glorot_uniform',
padding='same'))
model.add(LeakyReLU(0.1))
model.add(
ConvSN2D(128,
kernel_size=4,
strides=2,
kernel_initializer='glorot_uniform',
padding='same'))
model.add(LeakyReLU(0.1))
model.add(
ConvSN2D(256,
kernel_size=3,
strides=1,
kernel_initializer='glorot_uniform',
padding='same'))
model.add(LeakyReLU(0.1))
model.add(
ConvSN2D(256,
kernel_size=4,
strides=2,
kernel_initializer='glorot_uniform',
padding='same'))
model.add(LeakyReLU(0.1))
model.add(
ConvSN2D(512,
kernel_size=3,
strides=1,
kernel_initializer='glorot_uniform',
padding='same'))
model.add(LeakyReLU(0.1))
model.add(GlobalSumPooling2D())
model.add(DenseSN(1, kernel_initializer='glorot_uniform'))
else:
model = Sequential(name=name)
model.add(
Conv2D(64,
kernel_size=3,
strides=1,
kernel_initializer='glorot_uniform',
padding='same',
input_shape=(32, 32, 3)))
model.add(LeakyReLU(0.1))
model.add(
Conv2D(64,
kernel_size=4,
strides=2,
kernel_initializer='glorot_uniform',
padding='same'))
model.add(LeakyReLU(0.1))
model.add(
Conv2D(128,
kernel_size=3,
strides=1,
kernel_initializer='glorot_uniform',
padding='same'))
model.add(LeakyReLU(0.1))
model.add(
Conv2D(128,
kernel_size=4,
strides=2,
kernel_initializer='glorot_uniform',
padding='same'))
model.add(LeakyReLU(0.1))
model.add(
Conv2D(256,
kernel_size=3,
strides=1,
kernel_initializer='glorot_uniform',
padding='same'))
model.add(LeakyReLU(0.1))
model.add(
Conv2D(256,
kernel_size=4,
strides=2,
kernel_initializer='glorot_uniform',
padding='same'))
model.add(LeakyReLU(0.1))
model.add(
Conv2D(512,
kernel_size=3,
strides=1,
kernel_initializer='glorot_uniform',
padding='same'))
model.add(LeakyReLU(0.1))
model.add(GlobalSumPooling2D())
model.add(Dense(1, kernel_initializer='glorot_uniform'))
if plot:
plot_model(model, name + '.png', show_layer_names=True)
if summary:
print('Discriminator')
print('Spectral Normalization: {}'.format(spectral_normalization))
model.summary()
return model
def discriminator_aux(input_shape,
pc_code_dim=0,
bn_momentum=0.9,
bn_epsilon=0.00002):
inputs = Input(shape=input_shape, name='image_input')
channels = 64
x = ResBlock(input_shape=input_shape,
channels=channels,
sampling='down',
batch_normalization=False,
name='dis_resblock_0')(inputs)
d = input_shape[0] // 2
shape = (d, d, channels)
name = 'dis_resblock_1'
x = ResBlock(input_shape=shape,
channels=(channels * 2),
sampling='down',
batch_normalization=False,
name=name)(x)
channels *= 2
d = input_shape[0] // 4
shape = (d, d, channels)
name = 'dis_resblock_2'
x = ResBlock(input_shape=shape,
channels=(channels * 2),
sampling='down',
batch_normalization=False,
name=name)(x)
channels *= 2
d = input_shape[0] // 8
shape = (d, d, channels)
name = 'dis_resblock_3'
x = ResBlock(input_shape=shape,
channels=(channels * 2),
sampling='down',
batch_normalization=False,
name=name)(x)
channels *= 2
aux_layer = Flatten(name='aux_layer')(x)
d = input_shape[0] // 16
shape = (d, d, channels)
name = 'dis_resblock_4'
x = ResBlock(input_shape=shape,
channels=(channels * 2),
sampling='down',
batch_normalization=False,
name=name)(x)
channels *= 2
d = input_shape[0] // 32
shape = (d, d, channels)
name = 'dis_resblock_5'
x = ResBlock(input_shape=shape,
channels=channels,
sampling=None,
batch_normalization=False,
name=name)(x)
x = Activation('relu')(x)
x = GlobalSumPooling2D()(x)
preal = DenseSN(1,
activation='sigmoid',
kernel_initializer='glorot_uniform',
name='real_fake')(x)
aux_layer = DenseSN(256, activation='relu')(aux_layer)
reco_pc_code = DenseSN(1024, activation='relu')(aux_layer)
reco_pc_code = DenseSN(1024, activation='relu')(reco_pc_code)
reco_pc_code = DenseSN(1024, activation='relu')(reco_pc_code)
reco_pc_code = DenseSN(pc_code_dim,
activation='linear',
name='reco_pc_code')(reco_pc_code)
reco_elev_code = DenseSN(128, activation='relu')(aux_layer)
reco_elev_code = DenseSN(128, activation='relu')(reco_elev_code)
reco_elev_code = DenseSN(128, activation='relu')(reco_elev_code)
reco_elev_code = DenseSN(1,
activation='sigmoid',
kernel_initializer='glorot_uniform',
name='reco_elev_code')(reco_elev_code)
reco_azim_code = DenseSN(128, activation='relu')(aux_layer)
reco_azim_code = DenseSN(128, activation='relu')(reco_azim_code)
reco_azim_code = DenseSN(128, activation='relu')(reco_azim_code)
reco_azim_code = DenseSN(1,
activation='sigmoid',
kernel_initializer='glorot_uniform',
name='reco_azim_code')(reco_azim_code)
outputs = [preal, reco_pc_code, reco_elev_code, reco_azim_code]
return Model(inputs, outputs, name='discriminator')
def word_cnn(dataset, use_glove=False, sn=True):
filters = 250
kernel_size = 3
hidden_dims = 250
max_len = config.word_max_len[dataset]
num_classes = config.num_classes[dataset]
loss = config.loss[dataset]
activation = config.activation[dataset]
embedding_dims = config.wordCNN_embedding_dims[dataset]
num_words = config.num_words[dataset]
print('Build word_cnn model...')
model = Sequential()
if use_glove:
file_path = r'./glove.6B.{}d.txt'.format(str(embedding_dims))
pdb.set_trace()
get_embedding_index(file_path)
get_embedding_matrix(dataset, num_words, embedding_dims)
model.add(
Embedding( # Layer 0, Start
input_dim=num_words +
1, # Size to dictionary, has to be input + 1
output_dim=embedding_dims, # Dimensions to generate
weights=[embedding_matrix], # Initialize word weights
input_length=max_len,
name="embedding_layer",
trainable=False))
else:
# embedding_dims => 50
model.add(Embedding(num_words, embedding_dims, input_length=max_len))
model.add(Dropout(0.2))
if sn:
## with SN ##
model.add(Reshape((-1, 400, 50)))
model.add(
ConvSN2D(filters, (1, kernel_size),
strides=1,
padding='valid',
activation='relu'))
model.add(Reshape((398, 250)))
model.add(GlobalMaxPooling1D())
model.add(DenseSN(hidden_dims))
model.add(Dropout(0.2))
model.add(Activation('relu'))
model.add(DenseSN(num_classes))
model.add(Activation(activation))
else:
## without SN ##
model.add(
Conv1D(filters,
kernel_size,
padding='valid',
activation='relu',
strides=1))
model.add(GlobalMaxPooling1D())
model.add(Dense(hidden_dims))
model.add(Dropout(0.2))
model.add(Activation('relu'))
model.add(Dense(num_classes))
model.add(Activation(activation))
model.compile(loss=loss, optimizer='adam', metrics=['accuracy'])
return model
def discriminator(input_shape,
pc_code_dim=0,
spectral_normalization=True,
batch_normalization=False,
bn_momentum=0.9,
bn_epsilon=0.00002):
inputs = Input(shape=input_shape, name='image_input')
channels = 64
# conv2.1
x = ResBlockBottleneck(input_shape=input_shape,
channels=channels,
sampling='None',
batch_normalization=batch_normalization,
spectral_normalization=spectral_normalization,
name='dis_resblock_01')(inputs)
d = input_shape[0]
shape = (d, d, channels)
x = ResBlockBottleneck(input_shape=shape,
channels=channels,
sampling='None',
batch_normalization=batch_normalization,
spectral_normalization=spectral_normalization,
name='dis_resblock_02')(x)
x = ResBlockBottleneck(input_shape=shape,
channels=channels,
sampling='down',
batch_normalization=batch_normalization,
spectral_normalization=spectral_normalization,
name='dis_resblock_03')(x)
# conv2.2
d = input_shape[0] // 2
shape = (d, d, channels)
x = ResBlockBottleneck(input_shape=shape,
channels=channels * 2,
sampling='None',
batch_normalization=batch_normalization,
spectral_normalization=spectral_normalization,
name='dis_resblock_10')(x)
shape = (d, d, channels * 2)
x = ResBlockBottleneck(input_shape=shape,
channels=channels * 2,
sampling='None',
batch_normalization=batch_normalization,
spectral_normalization=spectral_normalization,
name='dis_resblock_11')(x)
x = ResBlockBottleneck(input_shape=shape,
channels=channels * 2,
sampling='None',
batch_normalization=batch_normalization,
spectral_normalization=spectral_normalization,
name='dis_resblock_12')(x)
x = ResBlockBottleneck(input_shape=shape,
channels=channels * 2,
sampling='down',
batch_normalization=batch_normalization,
spectral_normalization=spectral_normalization,
name='dis_resblock_13')(x)
channels *= 2
# conv2.3
d = input_shape[0] // 4
shape = (d, d, channels)
x = ResBlockBottleneck(input_shape=shape,
channels=channels * 2,
sampling='None',
batch_normalization=batch_normalization,
spectral_normalization=spectral_normalization,
name='dis_resblock_20')(x)
shape = (d, d, channels * 2)
x = ResBlockBottleneck(input_shape=shape,
channels=channels * 2,
sampling='None',
batch_normalization=batch_normalization,
spectral_normalization=spectral_normalization,
name='dis_resblock_21')(x)
x = ResBlockBottleneck(input_shape=shape,
channels=channels * 2,
sampling='None',
batch_normalization=batch_normalization,
spectral_normalization=spectral_normalization,
name='dis_resblock_22')(x)
x = ResBlockBottleneck(input_shape=shape,
channels=channels * 2,
sampling='None',
batch_normalization=batch_normalization,
spectral_normalization=spectral_normalization,
name='dis_resblock_23')(x)
x = ResBlockBottleneck(input_shape=shape,
channels=channels * 2,
sampling='None',
batch_normalization=batch_normalization,
spectral_normalization=spectral_normalization,
name='dis_resblock_24')(x)
x = ResBlockBottleneck(input_shape=shape,
channels=channels * 2,
sampling='down',
batch_normalization=batch_normalization,
spectral_normalization=spectral_normalization,
name='dis_resblock_25')(x)
channels *= 2
d = input_shape[0] // 8
shape = (d, d, channels)
x = ResBlockBottleneck(input_shape=shape,
channels=channels * 2,
sampling='None',
batch_normalization=batch_normalization,
spectral_normalization=spectral_normalization,
name='dis_resblock_30')(x)
shape = (d, d, channels * 2)
x = ResBlockBottleneck(input_shape=shape,
channels=channels * 2,
sampling='None',
batch_normalization=batch_normalization,
spectral_normalization=spectral_normalization,
name='dis_resblock_31')(x)
x = ResBlockBottleneck(input_shape=shape,
channels=channels * 2,
sampling='down',
batch_normalization=batch_normalization,
spectral_normalization=spectral_normalization,
name='dis_resblock_32')(x)
channels *= 2
# d = input_shape[0] // 2
# shape = (d, d, channels)
# name = 'dis_resblock_1'
# x = ResBlockBottleneck(input_shape=shape,
# channels=(channels * 2),
# sampling='down',
# batch_normalization=batch_normalization,
# spectral_normalization=spectral_normalization,
# name=name)(x)
# channels *= 2
#
# d = input_shape[0] // 4
# shape = (d, d, channels)
# name = 'dis_resblock_2'
# x = ResBlockBottleneck(input_shape=shape,
# channels=(channels * 2),
# sampling='down',
# batch_normalization=batch_normalization,
# spectral_normalization=spectral_normalization,
# name=name)(x)
# channels *= 2
#
# d = input_shape[0] // 8
# shape = (d, d, channels)
# name = 'dis_resblock_3'
# x = ResBlockBottleneck(input_shape=shape,
# channels=(channels * 2),
# sampling='down',
# batch_normalization=batch_normalization,
# spectral_normalization=spectral_normalization,
# name=name)(x)
# channels *= 2
aux_layer_pc = Flatten(name='aux_layer_pc')(x)
# d = input_shape[0] // 16
# shape = (d, d, channels)
# name = 'dis_resblock_4'
# x = ResBlockBottleneck(input_shape=shape,
# channels=(channels * 2),
# sampling='down',
# batch_normalization=batch_normalization,
# spectral_normalization=spectral_normalization,
# name=name)(x)
# channels *= 2
#
# d = input_shape[0] // 32
# shape = (d, d, channels)
# name = 'dis_resblock_5'
# x = ResBlockBottleneck(input_shape=shape,
# channels=channels,
# sampling=None,
# batch_normalization=batch_normalization,
# spectral_normalization=spectral_normalization,
# name=name)(x)
x = Activation('relu')(x)
x = GlobalSumPooling2D()(x)
preal = DenseSN(1,
activation='sigmoid',
kernel_initializer='glorot_uniform',
name='real_fake')(x)
aux_layer_pc = DenseSN(256, activation='relu')(aux_layer_pc)
reco_pc_code = DenseSN(1024, activation='relu')(aux_layer_pc)
reco_pc_code = DenseSN(1024, activation='relu')(reco_pc_code)
reco_pc_code = DenseSN(1024, activation='relu')(reco_pc_code)
reco_pc_code = DenseSN(pc_code_dim,
activation='linear',
name='reco_pc_code')(reco_pc_code)
reco_azim_code = DenseSN(1,
activation='sigmoid',
kernel_initializer='glorot_uniform',
name='reco_azim_code')(x)
reco_elev_code = DenseSN(1,
activation='sigmoid',
kernel_initializer='glorot_uniform',
name='reco_elev_code')(x)
outputs = [preal, reco_pc_code, reco_elev_code, reco_azim_code]
return Model(inputs, outputs, name='discriminator')