def __init__(self,
blocks=16,
hidden_units=128,
kernel_size=3,
*args,
**kwargs):
super(DCResNet, self).__init__(*args, **kwargs)
self.blocks = blocks
self.hidden_units = hidden_units
w_regularizer = keras.regularizers.l2(self.wd)
self.conv1 = keras.layers.Conv2D(hidden_units,
kernel_size,
activation=tf.nn.relu,
kernel_regularizer=w_regularizer,
bias_regularizer=w_regularizer,
padding="same")
self.concrete1 = bdl.ConcreteLayer(dropout_decay=self.dd)
self.residual_blocks = []
for _ in range(self.blocks):
self.residual_blocks.append(
ResidualBlock(hidden_units,
weight_decay=self.wd,
dropout_decay=self.dd))
self.concrete2 = bdl.ConcreteLayer(dropout_decay=self.dd)
w_regularizer = keras.regularizers.l2(self.wd)
self.conv2 = keras.layers.Conv2D(hidden_units,
kernel_size,
activation=tf.nn.relu,
kernel_regularizer=w_regularizer,
bias_regularizer=w_regularizer,
padding="same")
self.concrete3 = bdl.ConcreteLayer(dropout_decay=self.dd)
w_regularizer = keras.regularizers.l2(self.wd)
self.conv3 = keras.layers.Conv2D(self.output_bands * 2 + 1,
kernel_size,
activation=None,
kernel_regularizer=w_regularizer,
bias_regularizer=w_regularizer,
padding="same")
def __init__(self,
output_bands,
hidden_layers=(512, ) * 3,
*args,
**kwargs):
super(DCVDSR_New, self).__init__(output_bands, *args, **kwargs)
self.hidden_layers = hidden_layers
# w_regularizer = keras.regularizers.l2(self.wd)
self.conv1 = keras.layers.Conv2D(
self.hidden_layers[0],
4,
activation=tf.nn.relu,
# kernel_regularizer=w_regularizer,
# bias_regularizer=w_regularizer,
padding="same")
self.concrete_layers = []
self.conv_layers = []
for i, hunits in enumerate(self.hidden_layers):
self.concrete_layers.append(bdl.ConcreteLayer(self.dd))
# w_regularizer = keras.regularizers.l2(self.wd)
self.conv_layers.append([
keras.layers.Conv2D(hunits,
3,
activation=tf.nn.relu,
padding="same"),
keras.layers.Conv2D(hunits,
5,
activation=tf.nn.relu,
padding="same"),
keras.layers.Conv2D(hunits,
7,
activation=tf.nn.relu,
padding="same")
])
self.concrete_output = bdl.ConcreteLayer(self.dd)
# w_regularizer = keras.regularizers.l2(self.wd)
self.conv_output = keras.layers.Conv2D(
self.output_bands * 2 + 1,
3,
activation=None,
# kernel_regularizer=w_regularizer,
# bias_regularizer=w_regularizer,
padding="same")
def __init__(self,
hidden_units,
weight_decay=1e-4,
dropout_decay=1e-4,
bayesian=True):
super(ResidualBlock, self).__init__()
self.bayesian = bayesian
self.weight_decay = weight_decay
if bayesian:
self.concrete1 = bdl.ConcreteLayer(dropout_decay=dropout_decay)
w_regularizer = keras.regularizers.l2(weight_decay)
self.conv1 = keras.layers.Conv2D(hidden_units,
3,
activation=tf.nn.leaky_relu,
kernel_regularizer=w_regularizer,
bias_regularizer=w_regularizer,
padding="same")
if bayesian:
self.concrete2 = bdl.ConcreteLayer(dropout_decay=dropout_decay)
w_regularizer = keras.regularizers.l2(weight_decay)
self.conv2 = keras.layers.Conv2D(hidden_units,
3,
activation=None,
kernel_regularizer=w_regularizer,
bias_regularizer=w_regularizer,
padding="same")
if bayesian:
self.concrete3 = bdl.ConcreteLayer(dropout_decay=dropout_decay)
w_regularizer = keras.regularizers.l2(weight_decay)
self.conv3 = keras.layers.Conv2D(hidden_units,
3,
activation=tf.nn.leaky_relu,
kernel_regularizer=w_regularizer,
bias_regularizer=w_regularizer,
padding="same")
def __init__(self, filter_sizes, layer_sizes, *args, **kwargs):
super(DCCNN, self).__init__(*args, **kwargs)
self.filter_sizes = filter_sizes
self.layer_sizes = layer_sizes
self.concrete_layers = []
self.conv2d_layers = []
for i, k in enumerate(self.filter_sizes):
if i != 0:
concrete_layer = bdl.ConcreteLayer(dropout_decay=self.dd)
self.concrete_layers.append(concrete_layer)
if i == len(filter_sizes) - 1:
activation = None
else:
activation = tf.nn.relu
self.conv2d_layers.append(
keras.layers.Conv2D(self.layer_sizes[i],
k,
activation=activation,
padding='same'))