def __init__(
self,
n_mel_channels: int,
n_flows: int,
n_group: int,
n_early_every: int,
n_early_size: int,
n_wn_channels: int,
n_wn_layers: int,
wn_kernel_size: int,
):
"""
WaveGlow module
Args:
n_mel_channels (int): Number of mel channels to output.
n_flows (int): Number of flow layers
n_group (int): Number of groups to respace the inputs
n_early_every (int): Every n_early_every layers, n_early_size gets skip connected to the output
n_early_size (int): The size of the chunk to be skip connected
n_wn_channels (int): Number of channels for the non-invertible wavenet transformation
n_wn_layers (int): Number of layers for the non-invertible wavenet transformation
wn_kernel_size (int): Kernel size for the non-invertible wavenet transformation
"""
super().__init__()
self.upsample = torch.nn.ConvTranspose1d(n_mel_channels,
n_mel_channels,
1024,
stride=256)
self.n_mel_channels = n_mel_channels
assert n_group % 2 == 0
self.n_flows = n_flows
self.n_group = n_group
self.n_early_every = n_early_every
self.n_early_size = n_early_size
self.wavenet = torch.nn.ModuleList()
self.convinv = torch.nn.ModuleList()
self.mode = OperationMode.infer
n_half = n_group // 2
# Set up layers with the right sizes based on how many dimensions
# have been output already
n_remaining_channels = n_group
for k in range(n_flows):
if k % self.n_early_every == 0 and k > 0:
n_half = n_half - int(self.n_early_size / 2)
n_remaining_channels = n_remaining_channels - self.n_early_size
self.convinv.append(Invertible1x1Conv(n_remaining_channels))
self.wavenet.append(
WaveNet(
n_half,
n_mel_channels * n_group,
n_layers=n_wn_layers,
n_channels=n_wn_channels,
kernel_size=wn_kernel_size,
))
self.n_remaining_channels = n_remaining_channels
self.removed_weightnorm = False
def __init__(
self,
n_mel_channels: int,
n_flows: int,
n_group: int,
n_wn_channels: int,
n_wn_layers: int,
wn_kernel_size: int,
upsample_factor: int,
):
"""
UniGlow module
Args:
n_mel_channels (int): Number of mel channels to output.
n_flows (int): Number of flow layers
n_group (int): Number of groups to respace the inputs
n_early_every (int): Every n_early_every layers, n_early_size gets skip connected to the output
n_early_size (int): The size of the chunk to be skip connected
n_wn_channels (int): Number of channels for the non-invertible wavenet transformation
n_wn_layers (int): Number of layers for the non-invertible wavenet transformation
wn_kernel_size (int): Kernel size for the non-invertible wavenet transformation
"""
super().__init__()
assert n_group % 2 == 0
self.n_flows = n_flows
self.n_group = n_group
n_half = int(n_group / 2)
self.conv = Invertible1x1Conv(n_group)
self.wn = WaveNet(n_half, n_mel_channels, n_wn_layers, n_wn_channels, wn_kernel_size)
self.upsample_factor = upsample_factor