def __init__(self, tier: int, n_layers: int, hidden_size: int,
gmm_size: int, freq: int):
"""
Args:
tier (int): the tier that this module represents.
n_layers (int): number of layers this tier is composed of.
hidden_size (int): parameter for the hidden_state of the Delayed Stack Layers
gmm_size (int): number of mixture components of the GMM
freq (int): size of the frequency axis of the spectrogram to generate. See note in the
documentation of the file.
"""
super(Tier, self).__init__()
self.tier = tier
self.n_layers = n_layers
self.hidden_size = hidden_size
self.k = gmm_size
self.freq = freq
# Only the initial tier uses a centralized stack according to MelNet paper (Table 1)
self.has_central_stack = False
# Define layers of the tier
self.layers = nn.ModuleList([
ModuleWrapperDelayedStackLayer0(
DelayedStackLayer0(hidden_size=hidden_size,
has_central_stack=self.has_central_stack,
freq=freq,
is_conditioned=True,
hidden_size_condition=hidden_size * 4))
] + [
ModuleWrapperDelayedStackLayer(
DelayedStackLayer(layer=layer_idx,
hidden_size=hidden_size,
has_central_stack=self.has_central_stack,
freq=freq))
for layer_idx in range(1, n_layers)
])
# The Layer 0 of this tier (greater than first tier) is conditioned on the output of the
# feature extraction network. These conditioning features are the concatenation of the
# hidden state of 4 one dimensional RNN, that's why the hidden size of the condition in
# DelayedStackLayer0 has * 4
# Define feature extraction network
self.feature_extraction = ModuleWrapperFeatureExtraction(
FeatureExtractionLayer(hidden_size))
# Define dummy tensor to trick checkpointing
self.dummy_tensor = torch.ones(1,
dtype=torch.float32,
requires_grad=True)
# Linear transformation from final layer of the frequency-delayed stack to produce
# unconstrained parameters
self.W_theta = nn.Linear(in_features=hidden_size,
out_features=3 * self.k)
def __init__(self, tier: int, n_layers: int, hidden_size: int,
gmm_size: int, freq: int):
"""
Args:
tier (int): the tier that this module represents.
n_layers (int): number of layers this tier is composed of.
hidden_size (int): parameter for the hidden_state of the Delayed Stack Layers
gmm_size (int): number of mixture components of the GMM
freq (int): size of the frequency axis of the spectrogram to generate. See note in the
documentation of the file.
"""
super(Tier1, self).__init__()
self.tier = tier
self.n_layers = n_layers
self.hidden_size = hidden_size
self.k = gmm_size
# Only the initial tier uses a centralized stack according to MelNet paper (Table 1)
self.has_central_stack = True
# Define layers of the tier
self.layers = nn.ModuleList([
ModuleWrapperDelayedStackLayer0(
DelayedStackLayer0(hidden_size=hidden_size,
has_central_stack=self.has_central_stack,
freq=freq))
] + [
ModuleWrapperDelayedStackLayer(
DelayedStackLayer(layer=layer_idx,
hidden_size=hidden_size,
has_central_stack=self.has_central_stack,
freq=freq))
for layer_idx in range(1, n_layers)
])
# Define dummy tensor to trick checkpointing
self.dummy_tensor = torch.ones(1,
dtype=torch.float32,
requires_grad=True)
# Linear transformation from final layer of the frequency-delayed stack to produce
# unconstrained parameters
self.W_theta = nn.Linear(in_features=hidden_size,
out_features=3 * self.k)