def test_convtranpose1d(filters, kernel_size, strides, padding,
is_weight_norm):
fake_input_1d = tf.random.normal(shape=[4, 8000, 256], dtype=tf.float32)
conv1d_transpose = TFConvTranspose1d(filters=filters,
kernel_size=kernel_size,
strides=strides,
padding=padding,
is_weight_norm=is_weight_norm,
initializer_seed=42)
out = conv1d_transpose(fake_input_1d)
assert np.array_equal(tf.keras.backend.int_shape(out),
[4, 8000 * strides, filters])
def __init__(self, config, **kwargs):
super().__init__(**kwargs)
# check hyper parameter is valid or not
assert (
config.stacks
== len(config.stack_kernel_size)
== len(config.stack_dilation_rate)
)
# add initial layer
layers = []
layers += [
TFReflectionPad1d(
(config.kernel_size - 1) // 2,
padding_type=config.padding_type,
name="first_reflect_padding",
),
tf.keras.layers.Conv1D(
filters=config.filters,
kernel_size=config.kernel_size,
use_bias=config.use_bias,
),
]
for i, upsample_scale in enumerate(config.upsample_scales):
# add upsampling layer
layers += [
getattr(tf.keras.layers, config.nonlinear_activation)(
**config.nonlinear_activation_params
),
TFConvTranspose1d(
filters=config.filters // (2 ** (i + 1)),
kernel_size=upsample_scale * 2,
strides=upsample_scale,
padding="same",
is_weight_norm=config.is_weight_norm,
initializer_seed=config.initializer_seed,
name="conv_transpose_._{}".format(i),
),
]
# add residual stack layer
layers += [
TFMultiHifiResBlock(
list_resblock=[
TFHifiResBlock(
kernel_size=config.stack_kernel_size[j],
filters=config.filters // (2 ** (i + 1)),
dilation_rate=config.stack_dilation_rate[j],
use_bias=config.use_bias,
nonlinear_activation=config.nonlinear_activation,
nonlinear_activation_params=config.nonlinear_activation_params,
is_weight_norm=config.is_weight_norm,
initializer_seed=config.initializer_seed,
name="hifigan_resblock_._{}".format(j),
)
for j in range(config.stacks)
],
name="multi_hifigan_resblock_._{}".format(i),
)
]
# add final layer
layers += [
getattr(tf.keras.layers, config.nonlinear_activation)(
**config.nonlinear_activation_params
),
TFReflectionPad1d(
(config.kernel_size - 1) // 2,
padding_type=config.padding_type,
name="last_reflect_padding",
),
tf.keras.layers.Conv1D(
filters=config.out_channels,
kernel_size=config.kernel_size,
use_bias=config.use_bias,
dtype=tf.float32,
),
]
if config.use_final_nolinear_activation:
layers += [tf.keras.layers.Activation("tanh", dtype=tf.float32)]
if config.is_weight_norm is True:
self._apply_weightnorm(layers)
self.hifigan = tf.keras.models.Sequential(layers)
def __init__(self, config, encoder, **kwargs):
super().__init__(**kwargs)
# check hyper parameter is valid or not
assert config.filters >= np.prod(config.upsample_scales)
assert config.filters % (2 ** len(config.upsample_scales)) == 0
# add initial layer
self.encoder = encoder
gc_linear = []
layers = []
layers += [
TFReflectionPad1d(
(config.kernel_size - 1) // 2,
padding_type=config.padding_type,
name="first_reflect_padding",
),
tf.keras.layers.Conv1D(
filters=config.filters,
kernel_size=config.kernel_size,
use_bias=config.use_bias,
kernel_initializer=get_initializer(config.initializer_seed),
),
]
gc_linear += [
tf.keras.layers.Dense(
units=config.filters,
kernel_initializer=get_initializer(config.initializer_seed),
name='gc_start'
)
]
for i, upsample_scale in enumerate(config.upsample_scales):
# add upsampling layer
layers += [
getattr(tf.keras.layers, config.nonlinear_activation)(
**config.nonlinear_activation_params
),
TFConvTranspose1d(
filters=config.filters // (2 ** (i + 1)),
kernel_size=upsample_scale * 2,
strides=upsample_scale,
padding="same",
is_weight_norm=config.is_weight_norm,
initializer_seed=config.initializer_seed,
name="conv_transpose_._{}".format(i),
),
]
gc_linear += [
tf.keras.layers.Dense(
units=config.filters // (2 ** (i + 1)),
activation=tf.nn.tanh,
kernel_initializer=get_initializer(config.initializer_seed),
name='gc_%d_0'%(i)
)
]
# add residual stack layer
for j in range(config.stacks):
layers += [
TFResidualStack(
kernel_size=config.stack_kernel_size,
filters=config.filters // (2 ** (i + 1)),
dilation_rate=config.stack_kernel_size ** j,
use_bias=config.use_bias,
nonlinear_activation=config.nonlinear_activation,
nonlinear_activation_params=config.nonlinear_activation_params,
is_weight_norm=config.is_weight_norm,
initializer_seed=config.initializer_seed,
name="residual_stack_._{}._._{}".format(i, j),
)
]
gc_linear += [
tf.keras.layers.Dense(
units=config.filters // (2 ** (i + 1)),
activation=tf.nn.tanh,
kernel_initializer=get_initializer(config.initializer_seed),
name='gc_%d_%d'%(i, j+1)
)
]
# add final layer
gc_linear += [
tf.keras.layers.Dense(
units=config.out_channels,
kernel_initializer=get_initializer(config.initializer_seed),
name='gc_end'
)
]
layers += [
getattr(tf.keras.layers, config.nonlinear_activation)(
**config.nonlinear_activation_params
),
TFReflectionPad1d(
(config.kernel_size - 1) // 2,
padding_type=config.padding_type,
name="last_reflect_padding",
),
tf.keras.layers.Conv1D(
filters=config.out_channels,
kernel_size=config.kernel_size,
use_bias=config.use_bias,
kernel_initializer=get_initializer(config.initializer_seed),
dtype=tf.float32,
)
]
if config.use_final_nolinear_activation:
layers += [tf.keras.layers.Activation("tanh", dtype=tf.float32)]
if config.is_weight_norm is True:
self._apply_weightnorm(layers)
self.gc_linear = gc_linear
self.upsample = layers