def __init__(
self,
filters,
kernel_size,
strides,
padding,
is_weight_norm,
initializer_seed,
**kwargs
):
"""Initialize TFConvTranspose1d( module.
Args:
filters (int): Number of filters.
kernel_size (int): kernel size.
strides (int): Stride width.
padding (str): Padding type ("same" or "valid").
"""
super().__init__(**kwargs)
self.conv1d_transpose = tf.keras.layers.Conv2DTranspose(
filters=filters,
kernel_size=(kernel_size, 1),
strides=(strides, 1),
padding="same",
kernel_initializer=get_initializer(initializer_seed),
)
if is_weight_norm:
self.conv1d_transpose = WeightNormalization(self.conv1d_transpose)
def _apply_weightnorm(self, list_layers):
"""Try apply weightnorm for all layer in list_layers."""
for i in range(len(list_layers)):
try:
layer_name = list_layers[i].name.lower()
if "conv1d" in layer_name or "dense" in layer_name:
list_layers[i] = WeightNormalization(list_layers[i])
except Exception:
pass
def __init__(
self,
kernel_size,
filters,
dilation_rate,
use_bias,
nonlinear_activation,
nonlinear_activation_params,
is_weight_norm,
initializer_seed,
**kwargs
):
"""Initialize TFResidualStack module.
Args:
kernel_size (int): Kernel size.
filters (int): Number of filters.
dilation_rate (int): Dilation rate.
use_bias (bool): Whether to add bias parameter in convolution layers.
nonlinear_activation (str): Activation function module name.
nonlinear_activation_params (dict): Hyperparameters for activation function.
"""
super().__init__(**kwargs)
self.blocks = [
getattr(tf.keras.layers, nonlinear_activation)(
**nonlinear_activation_params
),
TFReflectionPad1d((kernel_size - 1) // 2 * dilation_rate),
tf.keras.layers.Conv1D(
filters=filters,
kernel_size=kernel_size,
dilation_rate=dilation_rate,
use_bias=use_bias,
kernel_initializer=get_initializer(initializer_seed),
),
getattr(tf.keras.layers, nonlinear_activation)(
**nonlinear_activation_params
),
tf.keras.layers.Conv1D(
filters=filters,
kernel_size=1,
use_bias=use_bias,
kernel_initializer=get_initializer(initializer_seed),
),
]
self.shortcut = tf.keras.layers.Conv1D(
filters=filters,
kernel_size=1,
use_bias=use_bias,
kernel_initializer=get_initializer(initializer_seed),
name="shortcut",
)
# apply weightnorm
if is_weight_norm:
self._apply_weightnorm(self.blocks)
self.shortcut = WeightNormalization(self.shortcut)
def __init__(
self,
period,
out_channels=1,
n_layers=5,
kernel_size=5,
strides=3,
filters=8,
filter_scales=4,
max_filters=1024,
nonlinear_activation="LeakyReLU",
nonlinear_activation_params={"alpha": 0.2},
initializer_seed=42,
is_weight_norm=False,
**kwargs
):
super().__init__(**kwargs)
self.period = period
self.out_filters = out_channels
self.convs = []
for i in range(n_layers):
self.convs.append(
tf.keras.layers.Conv2D(
filters=min(filters * (filter_scales ** (i + 1)), max_filters),
kernel_size=(kernel_size, 1),
strides=(strides, 1),
padding="same",
)
)
self.conv_post = tf.keras.layers.Conv2D(
filters=out_channels, kernel_size=(3, 1), padding="same",
)
self.activation = getattr(tf.keras.layers, nonlinear_activation)(
**nonlinear_activation_params
)
if is_weight_norm:
self._apply_weightnorm(self.convs)
self.conv_post = WeightNormalization(self.conv_post)