def __init__(self,
depthlen=288,
gen_features=8,
numparticle=6,
overscale=10.0,
expscale=False,
activation=tf.keras.activations.relu,
**kwargs):
super().__init__(**kwargs)
self.depthlen = depthlen
self.gen_features = gen_features
self.overscale = 10.0
self.expscale = expscale
self.labelmerger = utils.LabelMerger(numparticle=numparticle,
expscale=expscale)
self.activation = activation
if self.expscale:
last_activation = tf.keras.activations.elu
else:
last_activation = tf.keras.activations.sigmoid
self.layer1 = layers.Dense(512, activation=self.activation)
self.layer1_norm = layers.LayerNormalization(epsilon=1e-6)
self.layer2 = layers.Dense(1024, activation=self.activation)
self.layer2_norm = layers.LayerNormalization(epsilon=1e-6)
self.layer3 = layers.Dense(self.depthlen * self.gen_features,
activation=last_activation)
def __init__(self, numparticle=6,
expscale=False, epsilon=1e-25,
activation=tf.keras.activations.tanh, **kwargs):
super().__init__(**kwargs)
self._numparticle = numparticle
self.expscale = expscale
self.epsilon = epsilon
self.labelmerger = utils.LabelMerger(numparticle=numparticle,
expscale=expscale)
self.activation = activation
self.layer_label_flatten = layers.Flatten()
self.layer_data_flatten = layers.Flatten()
self.layer_dense1 = layers.Dense(1024, activation=self.activation)
self.layer_dense1_norm = layers.LayerNormalization(epsilon=1e-6)
self.layer_dense2 = layers.Dense(512, activation=self.activation)
self.layer_dense2_norm = layers.LayerNormalization(epsilon=1e-6)
self.layer_dense3 = layers.Dense(256, activation=self.activation)
self.layer_dense3_norm = layers.LayerNormalization(epsilon=1e-6)
self.layer_dense4 = layers.Dense(1, activation=None)
def __init__(self,
depthlen=288,
gen_features=8,
numparticle=6,
overscale=10.0,
expscale=False,
activation=tf.keras.activations.tanh,
**kwargs):
super().__init__(**kwargs)
self.depthlen = depthlen
self.gen_features = gen_features
self.overscale = 10.0
self.expscale = expscale
self.labelmerger = utils.LabelMerger(numparticle=numparticle,
expscale=expscale)
self.activation = activation
if self.expscale:
last_activation = tf.keras.activations.elu
else:
last_activation = tf.keras.activations.sigmoid
self._nheight = 4
self._nwidth = int(np.ceil(self.depthlen / (2 * 2 * 2 * 2 * 2)))
self._nfilter = 32
# dense label/noise combination
self.layer_combine1 = layers.Dense(512, activation=self.activation)
self.layer_combine2 = layers.Dense(self._nheight * self._nwidth *
self._nfilter,
activation=self.activation)
# 1st transpose then conv
self.layer_tc1 = layers.Conv2DTranspose(64, (2, 5),
strides=(1, 2),
padding="same",
activation=self.activation)
self.layer_rc1 = layers.Conv2D(64, (2, 5),
padding="same",
activation=self.activation)
# 2nd transpose then conv
self.layer_tc2 = layers.Conv2DTranspose(32, (2, 10),
strides=(1, 2),
padding="same",
activation=self.activation)
self.layer_rc2 = layers.Conv2D(32, (2, 10),
padding="same",
activation=self.activation)
# 3rd transpose then conv
self.layer_tc3 = layers.Conv2DTranspose(16, (2, 10),
strides=(1, 2),
padding="same",
activation=self.activation)
self.layer_rc3 = layers.Conv2D(16, (2, 10),
padding="same",
activation=self.activation)
# 4th transpose then conv
self.layer_tc4 = layers.Conv2DTranspose(32, (2, 5),
strides=(1, 2),
padding="same",
activation=self.activation)
self.layer_rc4 = layers.Conv2D(32, (2, 5),
padding="same",
activation=self.activation)
# 5th transpose then conv
self.layer_tc5 = layers.Conv2DTranspose(16, (2, 5),
strides=(1, 2),
padding="same",
activation=self.activation)
self.layer_rc5 = layers.Conv2D(self.gen_features, (self._nheight, 5),
padding="valid",
activation=last_activation)
def __init__(self,
numparticle=6,
expscale=False,
epsilon=1e-25,
activation=layers.LeakyReLU(alpha=0.01),
**kwargs):
super().__init__(**kwargs)
self._numparticle = numparticle
self.expscale = expscale
self.epsilon = epsilon
self.labelmerger = utils.LabelMerger(numparticle=numparticle,
expscale=expscale)
self.activation = activation
# transform label to data format
self.layer_combine_flatten = layers.Flatten()
self.layer_combine1 = layers.Dense(256, activation=self.activation)
self.layer_combine1_norm = layers.LayerNormalization(epsilon=1e-6)
# downstream
self.layer_conv1 = layers.Conv1D(64,
16,
padding="same",
activation=self.activation)
self.layer_conv2 = layers.Conv1D(56,
14,
padding="same",
activation=self.activation)
self.layer_conv3 = layers.Conv1D(48,
12,
padding="same",
activation=self.activation)
self.layer_conv4 = layers.Conv1D(40,
10,
padding="same",
activation=self.activation)
self.layer_conv5 = layers.Conv1D(32,
8,
padding="same",
activation=self.activation)
self.layer_conv6 = layers.Conv1D(24,
6,
padding="same",
activation=self.activation)
self.layer_conv7 = layers.Conv1D(16,
4,
padding="same",
activation=self.activation)
self.layer_conv1_norm = tfa.layers.InstanceNormalization(epsilon=1e-6)
self.layer_conv2_norm = tfa.layers.InstanceNormalization(epsilon=1e-6)
self.layer_conv3_norm = tfa.layers.InstanceNormalization(epsilon=1e-6)
self.layer_conv4_norm = tfa.layers.InstanceNormalization(epsilon=1e-6)
self.layer_conv5_norm = tfa.layers.InstanceNormalization(epsilon=1e-6)
self.layer_conv6_norm = tfa.layers.InstanceNormalization(epsilon=1e-6)
self.layer_conv7_norm = tfa.layers.InstanceNormalization(epsilon=1e-6)
self.layer_end_flatten = layers.Flatten()
self.layer_end_dense1 = layers.Dense(256, activation=self.activation)
self.layer_end_dense1_norm = layers.LayerNormalization(epsilon=1e-6)
self.layer_end_dense2 = layers.Dense(1, activation=None)