File size: 2,169 Bytes
9b2107c |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 |
import torch
import torch.nn as nn # pylint: disable=consider-using-from-import
from TTS.tts.layers.delightful_tts.conv_layers import ConvTransposed
class PhonemeProsodyPredictor(nn.Module):
"""Non-parallel Prosody Predictor inspired by: https://arxiv.org/pdf/2102.00851.pdf
It consists of 2 layers of 1D convolutions each followed by a relu activation, layer norm
and dropout, then finally a linear layer.
Args:
hidden_size (int): Size of hidden channels.
kernel_size (int): Kernel size for the conv layers.
dropout: (float): Probability of dropout.
bottleneck_size (int): bottleneck size for last linear layer.
lrelu_slope (float): Slope of the leaky relu.
"""
def __init__(
self,
hidden_size: int,
kernel_size: int,
dropout: float,
bottleneck_size: int,
lrelu_slope: float,
):
super().__init__()
self.d_model = hidden_size
self.layers = nn.ModuleList(
[
ConvTransposed(
self.d_model,
self.d_model,
kernel_size=kernel_size,
padding=(kernel_size - 1) // 2,
),
nn.LeakyReLU(lrelu_slope),
nn.LayerNorm(self.d_model),
nn.Dropout(dropout),
ConvTransposed(
self.d_model,
self.d_model,
kernel_size=kernel_size,
padding=(kernel_size - 1) // 2,
),
nn.LeakyReLU(lrelu_slope),
nn.LayerNorm(self.d_model),
nn.Dropout(dropout),
]
)
self.predictor_bottleneck = nn.Linear(self.d_model, bottleneck_size)
def forward(self, x: torch.Tensor, mask: torch.Tensor) -> torch.Tensor:
"""
Shapes:
x: :math: `[B, T, D]`
mask: :math: `[B, T]`
"""
mask = mask.unsqueeze(2)
for layer in self.layers:
x = layer(x)
x = x.masked_fill(mask, 0.0)
x = self.predictor_bottleneck(x)
return x
|