def model_SCNN_custom_v2_ps3(input_shape, classes):
kernel_size = (3, 3)
pool_size = (3, 3)
model = Sequential()
model.add(
SConv2D(filters=32,
kernel_size=kernel_size,
padding='SAME',
activation=tf.nn.relu,
input_shape=input_shape))
model.add(
SConv2D(filters=32,
kernel_size=kernel_size,
strides=(2, 2),
padding='SAME',
activation=tf.nn.relu))
model.add(Dropout(rate=0.25))
model.add(
SConv2D(filters=64,
kernel_size=kernel_size,
padding='SAME',
activation=tf.nn.relu))
model.add(SMaxPool2D(pool_size=pool_size, padding='SAME'))
model.add(Dropout(rate=0.25))
model.add(Flatten())
model.add(Dense(units=128, activation=tf.nn.relu))
model.add(Dropout(rate=0.5))
model.add(Dense(units=classes, activation=tf.nn.softmax))
return model
def resnet_layer(inputs,
num_filters=16,
filter_size=1.0,
kernel_size=3,
strides=1,
activation='relu',
batch_normalization=True,
conv_first=True,
mode='baseline'):
'''2D Convolution-Batch Normalization-Activation stack builder
# Arguments
inputs (tensor): input tensor from input image or previous layer
num_filters (int): Conv2D number of filters
filter_size (float): filter size factor (convolutional layers)
kernel_size (int): Conv2D square kernel dimensions
strides (int): Conv2D square stride dimensions
activation (string): activation name
batch_normalization (bool): whether to include batch normalization
conv_first (bool): conv-bn-activation (True) or
bn-activation-conv (False)
# Returns
x (tensor): tensor as input to the next layer
'''
num_filters = int(filter_size * num_filters)
if mode == 'baseline':
conv = Conv2D(num_filters,
kernel_size=kernel_size,
strides=strides,
padding='same',
kernel_initializer='he_normal',
kernel_regularizer=l2(1e-4))
elif mode == 'S-ResNet':
conv = SConv2D(filters=num_filters,
kernel_size=kernel_size,
strides=strides,
padding='SAME')
elif mode == 'H-ResNet':
conv = HConv2D(filters=num_filters,
kernel_size=kernel_size,
strides=strides,
padding='SAME')
x = inputs
if conv_first:
x = conv(x)
if batch_normalization:
x = BatchNormalization()(x)
if activation is not None:
x = Activation(activation)(x)
else:
if batch_normalization:
x = BatchNormalization()(x)
if activation is not None:
x = Activation(activation)(x)
x = conv(x)
return x
def model_CNN_SConv2D(input_shape, classes, kernel_size, pool_size):
model = Sequential()
model.add(
SConv2D(filters=32,
kernel_size=kernel_size,
padding='SAME',
activation=tf.nn.relu,
input_shape=input_shape))
model.add(MaxPool2D(pool_size=pool_size, padding='SAME'))
model.add(Dropout(rate=0.25))
model.add(
SConv2D(filters=64,
kernel_size=kernel_size,
padding='SAME',
activation=tf.nn.relu))
model.add(MaxPool2D(pool_size=pool_size, padding='SAME'))
model.add(Dropout(rate=0.25))
model.add(Flatten())
model.add(Dense(units=128, activation=tf.nn.relu))
model.add(Dropout(rate=0.5))
model.add(Dense(units=classes, activation=tf.nn.softmax))
return model
def build_generator(self):
model = Sequential()
# https://github.com/eriklindernoren/Keras-GAN/blob/master/acgan/acgan.py
# https://github.com/keras-team/keras/blob/master/examples/mnist_acgan.py
model.add(
Dense(units=128 * 8 * 8,
activation='relu',
input_dim=self.latent_dim))
model.add(Reshape(target_shape=(8, 8, 128)))
if self.mode == 'baseline':
# model.add(UpSampling2D())
# model.add(Conv2D(filters=128, kernel_size=3, padding='same'))
model.add(
Conv2DTranspose(filters=128,
kernel_size=3,
strides=2,
padding='same'))
elif self.mode == 'S-ACGAN':
# model.add(SSampling2D())
# model.add(SConv2D(filters=128, kernel_size=3, padding='SAME'))
model.add(
SConv2DTranspose(filters=128,
kernel_size=3,
strides=2,
padding='SAME'))
elif self.mode == 'H-ACGAN':
# model.add(HSampling2D())
# model.add(HConv2D(filters=128, kernel_size=3, padding='SAME'))
model.add(
HConv2DTranspose(filters=128,
kernel_size=3,
strides=2,
padding='SAME'))
model.add(Activation('relu'))
model.add(BatchNormalization())
if self.mode == 'baseline':
# model.add(UpSampling2D())
# model.add(Conv2D(filters=64, kernel_size=3, padding='same'))
model.add(
Conv2DTranspose(filters=64,
kernel_size=3,
strides=2,
padding='same'))
elif self.mode == 'S-ACGAN':
# model.add(SSampling2D())
# model.add(SConv2D(filters=64, kernel_size=3, padding='SAME'))
model.add(
SConv2DTranspose(filters=64,
kernel_size=3,
strides=2,
padding='SAME'))
elif self.mode == 'H-ACGAN':
# model.add(HSampling2D())
# model.add(HConv2D(filters=64, kernel_size=3, padding='SAME'))
model.add(
HConv2DTranspose(filters=64,
kernel_size=3,
strides=2,
padding='SAME'))
model.add(Activation('relu'))
model.add(BatchNormalization())
if self.mode == 'baseline':
model.add(
Conv2D(filters=self.channels, kernel_size=3, padding='same'))
elif self.mode == 'S-ACGAN':
model.add(
SConv2D(filters=self.channels, kernel_size=3, padding='SAME'))
elif self.mode == 'H-ACGAN':
model.add(
HConv2D(filters=self.channels, kernel_size=3, padding='SAME'))
model.add(Activation('tanh'))
noise = Input(shape=(self.latent_dim, ))
label = Input(shape=(1, ), dtype='int32')
label_embedding = Flatten()(Embedding(self.classes,
self.latent_dim)(label))
model_input = multiply([noise, label_embedding])
img = model(model_input)
return (Model([noise, label], img), model)
def build_discriminator(self):
model = Sequential()
# https://github.com/keras-team/keras/blob/master/examples/mnist_acgan.py
if self.mode == 'baseline':
model.add(
Conv2D(filters=32,
kernel_size=3,
strides=2,
padding='same',
input_shape=self.input_shape))
elif self.mode == 'S-ACGAN':
model.add(
SConv2D(filters=32,
kernel_size=3,
strides=2,
padding='SAME',
input_shape=self.input_shape))
elif self.mode == 'H-ACGAN':
model.add(
HConv2D(filters=32,
kernel_size=3,
strides=2,
padding='SAME',
input_shape=self.input_shape))
model.add(LeakyReLU(alpha=0.2))
model.add(Dropout(rate=0.3))
if self.mode == 'baseline':
model.add(Conv2D(filters=64, kernel_size=3, padding='same'))
elif self.mode == 'S-ACGAN':
model.add(SConv2D(filters=64, kernel_size=3, padding='SAME'))
elif self.mode == 'H-ACGAN':
model.add(HConv2D(filters=64, kernel_size=3, padding='SAME'))
model.add(LeakyReLU(alpha=0.2))
model.add(Dropout(rate=0.3))
if self.mode == 'baseline':
model.add(
Conv2D(filters=128, kernel_size=3, strides=2, padding='same'))
elif self.mode == 'S-ACGAN':
model.add(
SConv2D(filters=128, kernel_size=3, strides=2, padding='SAME'))
elif self.mode == 'H-ACGAN':
model.add(
HConv2D(filters=128, kernel_size=3, strides=2, padding='SAME'))
model.add(LeakyReLU(alpha=0.2))
model.add(Dropout(rate=0.3))
if self.mode == 'baseline':
model.add(Conv2D(filters=256, kernel_size=3, padding='same'))
elif self.mode == 'S-ACGAN':
model.add(SConv2D(filters=256, kernel_size=3, padding='SAME'))
elif self.mode == 'H-ACGAN':
model.add(HConv2D(filters=256, kernel_size=3, padding='SAME'))
model.add(LeakyReLU(alpha=0.2))
model.add(Dropout(rate=0.3))
model.add(Flatten())
img = Input(shape=self.input_shape)
features = model(img)
validity = Dense(units=1, activation='sigmoid')(features)
label = Dense(units=self.classes, activation='softmax')(features)
return (Model(img, [validity, label]), model)