def CVAE(input_shape, latent_dim=3, lr=0.001):
image_size = input_shape[:-1]
channels = input_shape[-1]
conv_layers = 4
feature_maps = [32, 64, 64, 32]
filter_shapes = [(3, 3), (3, 3), (3, 3), (3, 3)]
strides = [(1, 1), (2, 2), (1, 1), (1, 1)]
dense_layers = 1
dense_neurons = [128]
dense_dropouts = [0.5]
feature_maps = feature_maps[0:conv_layers]
filter_shapes = filter_shapes[0:conv_layers]
strides = strides[0:conv_layers]
autoencoder = conv_variational_autoencoder(image_size,
channels,
conv_layers,
feature_maps,
filter_shapes,
strides,
dense_layers,
dense_neurons,
dense_dropouts,
latent_dim,
lr=lr)
autoencoder.model.summary()
return autoencoder
def compile(self):
"""
Builds autoencoder.
"""
print("Building convolutional variational autoencoder")
# set up parameter
self.feature_maps = self.feature_maps[0:self.conv_layers]
self.filter_shapes = self.filter_shapes[0:self.conv_layers]
self.strides = self.strides[0:self.conv_layers]
self.autoencoder = conv_variational_autoencoder(
self.image_size, self.channels, self.conv_layers,
self.feature_maps, self.filter_shapes, self.strides,
self.dense_layers, self.dense_neurons, self.dense_dropouts,
self.latent_dim)
tensor_shape = (1, image_size[0], image_size[1])
else:
tensor_shape = (image_size[0], image_size[1], 1)
X_train = X_train.reshape((X_train.shape[0], ) + tensor_shape)
X_test = X_test.reshape((X_test.shape[0], ) + tensor_shape)
print "reshaped data:", "train:", np.shape(X_train), "test:", np.shape(X_test)
# build autoencoder;
print "building variational autoencoder"
# set up parameter;
feature_maps = feature_maps[0:conv_layers]
filter_shapes = filter_shapes[0:conv_layers]
strides = strides[0:conv_layers]
autoencoder = conv_variational_autoencoder(image_size, channels, conv_layers,
feature_maps, filter_shapes,
strides, dense_layers,
dense_neurons, dense_dropouts,
latent_dim)
# train, save & load;
for i in range(nb_start, nb_end):
if i == 0:
print("skipping - no previous saved file to load")
# load model;
else:
autoencoder.load("./model/model_%i" % i)
# train model;
autoencoder.train(X_train[0:],
batch_size,
epochs=epochs,
validation_data=(X_test[0:], X_test[0:]),