def create_model(self, input_shape=(1, 32, 32), dense_dim=1000, dy=10, nb_filters=[64, 128], kernel_size=(3, 3), pool_size=(2, 2),
dropout=0.5, bn=True, output_activation='softmax', opt='adam'):
"""
Create DRCN model: convnet model followed by conv. autoencoder
Args:
_input (Tensor) : input layer
dense_dim (int) : dimensionality of the final dense layers
dy (int) : output dimensionality
nb_filter (list) : list of #Conv2D filters
kernel_size (tuple) : Conv2D kernel size
pool_size (tuple) : MaxPool kernel size
dropout (float) : dropout rate
bn (boolean) : batch normalization mode
output_activation (string) : act. function for output layer
opt (string) : optimizer
"""
[d1, d2, c] = input_shape
if opt == 'adam':
opt = Adam(lr=3e-4)
elif opt == 'rmsprop':
opt = RMSprop(lr=1e-4)
_input = Input(shape=input_shape)
# Create ConvNet
self.create_convnet(_input, dense_dim=dense_dim, dy=dy, nb_filters=nb_filters,
kernel_size=kernel_size, pool_size=pool_size, dropout=dropout,
bn=bn, output_activation=output_activation, opt=opt)
# Create ConvAE, encoder functions are shared with ConvNet
_h = _input
# Reconstruct Conv2D layers
for i, nf in enumerate(nb_filters):
_h = self.enc_functions[i](_h)
_h = Activation('relu')(_h)
if i < 2:
_h = MaxPooling2D(pool_size=pool_size, padding='same')(_h)
[_, wflat, hflat, cflat] = _h.get_shape().as_list()
_h = Flatten()(_h)
# Dense layers
for i in range(len(nb_filters), len(self.enc_functions)):
_h = self.enc_functions[i](_h)
_h = Activation('relu')(_h)
# Decoder
_h = Dense(dense_dim)(_h)
_h = Activation('relu')(_h)
_xdec = Dense(wflat*hflat*cflat)(_h)
_xdec = Activation('relu')(_xdec)
_xdec = Reshape((wflat, hflat, nb_filters[-1]))(_xdec)
i = 0
for nf in reversed(nb_filters):
_xdec = Conv2D(nf, kernel_size, padding='same')(_xdec)
_xdec = Activation('relu')(_xdec)
if i > 0:
_xdec = UpSampling2D(size=pool_size)(_xdec)
i += 1
_xdec = Conv2D(c, kernel_size, padding='same', activation=clip_relu)(_xdec)
self.convae_model = Model(input=_input, output=_xdec)
self.convae_model.compile(loss='mse', optimizer=opt)
print(self.convae_model.summary())
act3 = Activation('relu')(conv3)
#conv4 = Convolution2D(64, 3, 3,border_mode='same')(act3)
#act4 = Activation('relu')(conv4)
pool2 = MaxPooling2D(pool_size=(2, 2))(act3)
conv5 = Conv2D(8, (3, 3), padding='same')(pool2)
act5 = Activation('relu')(conv5)
pool3 = MaxPooling2D(pool_size=(2, 2))(act5)
conv6 = Conv2D(8, (3, 3), padding='same')(pool3)
act6 = Activation('relu')(conv6)
pool4 = MaxPooling2D(pool_size=(2, 2))(act6)
conv7 = Conv2D(16, (3, 3), padding='same')(pool4)
act7 = Activation('tanh')(conv7)
#pool5 = MaxPooling2D(pool_size=(2, 2))(act7)
#conv8 = Convolution2D(32, 3, 3,border_mode='same')(pool5)
#act8 = Activation('relu')(conv8)
print(act7.get_shape())
encoder = Model(input=[input_lr], output=[act7])
encoder.compile(optimizer=adam, loss='mse', metrics=['accuracy'])
code_lr = Input(shape=(8, 8, 16), dtype='float32', name='decode_input')
print(code_lr.get_shape())
#dgaussian = GaussianNoise(0.1)(code_lr)
#deconv0 = Convolution2D(512, 3, 3, border_mode='same')(code_lr)
#deact0 = Activation('relu')(deconv0)
#dfc1 = Dense(32*4*4, activation='relu')(code_lr)
#dfc2 = Dense(16*8*8, activation='relu')(dgaussian)
#rshp1 = Reshape((16,8,8))(dfc2)
#up0 = UpSampling2D(size=(2,2))(rshp1)
#deconv1 = Convolution2D(32, 3, 3, border_mode='same')(up0)
#deact1 = Activation('relu')(deconv1)
def create_model(self,
input_shape=(1, 32, 32),
dense_dim=1000,
dy=10,
nb_filters=[64, 128],
kernel_size=(3, 3),
pool_size=(2, 2),
dropout=0.5,
bn=True,
output_activation='softmax',
opt='adam'):
"""
Create DRCN model: convnet model followed by conv. autoencoder
Args:
_input (Tensor) : input layer
dense_dim (int) : dimensionality of the final dense layers
dy (int) : output dimensionality
nb_filter (list) : list of #Conv2D filters
kernel_size (tuple) : Conv2D kernel size
pool_size (tuple) : MaxPool kernel size
dropout (float) : dropout rate
bn (boolean) : batch normalization mode
output_activation (string) : act. function for output layer
opt (string) : optimizer
"""
[d1, d2, c] = input_shape
if opt == 'adam':
opt = Adam(lr=3e-4)
elif opt == 'rmsprop':
opt = RMSprop(lr=1e-4)
_input = Input(shape=input_shape)
# Create ConvNet
self.create_convnet(_input,
dense_dim=dense_dim,
dy=dy,
nb_filters=nb_filters,
kernel_size=kernel_size,
pool_size=pool_size,
dropout=dropout,
bn=bn,
output_activation=output_activation,
opt=opt)
# Create ConvAE, encoder functions are shared with ConvNet
_h = _input
# Reconstruct Conv2D layers
for i, nf in enumerate(nb_filters):
_h = self.enc_functions[i](_h)
_h = Activation('relu')(_h)
if i < 2:
_h = MaxPooling2D(pool_size=pool_size, padding='same')(_h)
[_, wflat, hflat, cflat] = _h.get_shape().as_list()
_h = Flatten()(_h)
# Dense layers
for i in range(len(nb_filters), len(self.enc_functions)):
_h = self.enc_functions[i](_h)
_h = Activation('relu')(_h)
# Decoder
_h = Dense(dense_dim)(_h)
_h = Activation('relu')(_h)
_xdec = Dense(wflat * hflat * cflat)(_h)
_xdec = Activation('relu')(_xdec)
_xdec = Reshape((wflat, hflat, nb_filters[-1]))(_xdec)
i = 0
for nf in reversed(nb_filters):
_xdec = Conv2D(nf, kernel_size, padding='same')(_xdec)
_xdec = Activation('relu')(_xdec)
if i > 0:
_xdec = UpSampling2D(size=pool_size)(_xdec)
i += 1
_xdec = Conv2D(c, kernel_size, padding='same',
activation=clip_relu)(_xdec)
self.convae_model = Model(input=_input, output=_xdec)
self.convae_model.compile(loss='mse', optimizer=opt)
print(self.convae_model.summary())
