File size: 6,085 Bytes
705a8fd | 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 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 | import torch
import torch.nn as nn
import torch.nn.functional as F
from torch.nn.utils import weight_norm
from vector_quantize_pytorch import ResidualVQ
class CausalConv1d(nn.Conv1d):
def __init__(self, *args, **kwargs):
super().__init__(*args, **kwargs)
self.causal_padding = self.dilation[0] * (self.kernel_size[0] - 1)
def forward(self, x):
return self._conv_forward(F.pad(x, [self.causal_padding, 0]), self.weight, self.bias)
class CausalConvTranspose1d(nn.ConvTranspose1d):
def __init__(self, *args, **kwargs):
super().__init__(*args, **kwargs)
self.causal_padding = self.dilation[0] * (self.kernel_size[0] - 1) + self.output_padding[0] + 1 - self.stride[0]
def forward(self, x, output_size=None):
if self.padding_mode != 'zeros':
raise ValueError('Only `zeros` padding mode is supported for ConvTranspose1d')
assert isinstance(self.padding, tuple)
output_padding = self._output_padding(
x, output_size, self.stride, self.padding, self.kernel_size, self.dilation)
return F.conv_transpose1d(
x, self.weight, self.bias, self.stride, self.padding,
output_padding, self.groups, self.dilation)[...,:-self.causal_padding]
class ResidualUnit(nn.Module):
def __init__(self, in_channels, out_channels, dilation):
super().__init__()
self.dilation = dilation
self.layers = nn.Sequential(
CausalConv1d(in_channels=in_channels, out_channels=out_channels,
kernel_size=7, dilation=dilation),
nn.ELU(),
nn.Conv1d(in_channels=in_channels, out_channels=out_channels,
kernel_size=1)
)
def forward(self, x):
return x + self.layers(x)
class EncoderBlock(nn.Module):
def __init__(self, out_channels, stride):
super().__init__()
self.layers = nn.Sequential(
ResidualUnit(in_channels=out_channels//2,
out_channels=out_channels//2, dilation=1),
nn.ELU(),
ResidualUnit(in_channels=out_channels//2,
out_channels=out_channels//2, dilation=3),
nn.ELU(),
ResidualUnit(in_channels=out_channels//2,
out_channels=out_channels//2, dilation=9),
nn.ELU(),
CausalConv1d(in_channels=out_channels//2, out_channels=out_channels,
kernel_size=2*stride, stride=stride)
)
def forward(self, x):
return self.layers(x)
class DecoderBlock(nn.Module):
def __init__(self, out_channels, stride):
super().__init__()
self.layers = nn.Sequential(
CausalConvTranspose1d(in_channels=2*out_channels,
out_channels=out_channels,
kernel_size=2*stride, stride=stride),
nn.ELU(),
ResidualUnit(in_channels=out_channels, out_channels=out_channels,
dilation=1),
nn.ELU(),
ResidualUnit(in_channels=out_channels, out_channels=out_channels,
dilation=3),
nn.ELU(),
ResidualUnit(in_channels=out_channels, out_channels=out_channels,
dilation=9),
)
def forward(self, x):
return self.layers(x)
class Encoder(nn.Module):
def __init__(self, C, D):
super().__init__()
self.layers = nn.Sequential(
CausalConv1d(in_channels=2, out_channels=C, kernel_size=7),
nn.ELU(),
EncoderBlock(out_channels=2*C, stride=2),
nn.ELU(),
EncoderBlock(out_channels=4*C, stride=4),
nn.ELU(),
EncoderBlock(out_channels=8*C, stride=5),
nn.ELU(),
# EncoderBlock(out_channels=16*C, stride=8),
# nn.ELU(),
# CausalConv1d(in_channels=16*C, out_channels=D, kernel_size=3)
CausalConv1d(in_channels=8*C, out_channels=D, kernel_size=3)
)
def forward(self, x):
return self.layers(x)
class Decoder(nn.Module):
def __init__(self, C, D):
super().__init__()
self.layers = nn.Sequential(
CausalConv1d(in_channels=D, out_channels=8*C, kernel_size=7),
# CausalConv1d(in_channels=D, out_channels=16*C, kernel_size=7),
# nn.ELU(),
# DecoderBlock(out_channels=8*C, stride=8),
nn.ELU(),
DecoderBlock(out_channels=4*C, stride=5),
nn.ELU(),
DecoderBlock(out_channels=2*C, stride=4),
nn.ELU(),
DecoderBlock(out_channels=C, stride=2),
nn.ELU(),
CausalConv1d(in_channels=C, out_channels=2, kernel_size=7)
)
def forward(self, x):
return self.layers(x)
class SoundStream(nn.Module):
def __init__(self, C, D, n_q, codebook_size):
super().__init__()
self.encoder = Encoder(C=C, D=D)
self.quantizer = ResidualVQ(
num_quantizers=n_q, dim=D, codebook_size=codebook_size,
kmeans_init=True, kmeans_iters=100, threshold_ema_dead_code=2
)
self.decoder = Decoder(C=C, D=D)
@staticmethod
def pad_to_multiple(x, multiple):
"""
x: [B, C, T]
multiple: int, e.g., 320
return: padded_x, original_length
"""
B, C, T = x.shape
target_len = ((T + multiple - 1) // multiple) * multiple
pad_len = target_len - T
padded_x = F.pad(x, (0, pad_len), mode='reflect')
return padded_x, T
@staticmethod
def crop_to_length(x, original_length):
return x[..., :original_length]
def forward(self, x):
e = self.encoder(x) # [B, D, T']
e = e.permute(0, 2, 1) # → [B, T', D]
quantized, _, _ = self.quantizer(e)
quantized = quantized.permute(0, 2, 1) # → [B, D, T']
o = self.decoder(quantized) # → [B, 2, T_padded]
return o |