File size: 4,345 Bytes
4f175c5 | 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 | from typing import List, Tuple
import torch
import torch.nn as nn
from models.scnet_unofficial.utils import create_intervals
class Downsample(nn.Module):
def __init__(
self,
input_dim: int,
output_dim: int,
stride: int,
):
super().__init__()
self.conv = nn.Conv2d(input_dim, output_dim, 1, (stride, 1))
def forward(self, x: torch.Tensor) -> torch.Tensor:
return self.conv(x)
class ConvolutionModule(nn.Module):
def __init__(
self,
input_dim: int,
hidden_dim: int,
kernel_sizes: List[int],
bias: bool = False,
) -> None:
super().__init__()
self.sequential = nn.Sequential(
nn.GroupNorm(num_groups=1, num_channels=input_dim),
nn.Conv1d(
input_dim,
2 * hidden_dim,
kernel_sizes[0],
stride=1,
padding=(kernel_sizes[0] - 1) // 2,
bias=bias,
),
nn.GLU(dim=1),
nn.Conv1d(
hidden_dim,
hidden_dim,
kernel_sizes[1],
stride=1,
padding=(kernel_sizes[1] - 1) // 2,
groups=hidden_dim,
bias=bias,
),
nn.GroupNorm(num_groups=1, num_channels=hidden_dim),
nn.SiLU(),
nn.Conv1d(
hidden_dim,
input_dim,
kernel_sizes[2],
stride=1,
padding=(kernel_sizes[2] - 1) // 2,
bias=bias,
),
)
def forward(self, x: torch.Tensor) -> torch.Tensor:
x = x.transpose(1, 2)
x = x + self.sequential(x)
x = x.transpose(1, 2)
return x
class SDLayer(nn.Module):
def __init__(
self,
subband_interval: Tuple[float, float],
input_dim: int,
output_dim: int,
downsample_stride: int,
n_conv_modules: int,
kernel_sizes: List[int],
bias: bool = True,
):
super().__init__()
self.subband_interval = subband_interval
self.downsample = Downsample(input_dim, output_dim, downsample_stride)
self.activation = nn.GELU()
conv_modules = [
ConvolutionModule(
input_dim=output_dim,
hidden_dim=output_dim // 4,
kernel_sizes=kernel_sizes,
bias=bias,
)
for _ in range(n_conv_modules)
]
self.conv_modules = nn.Sequential(*conv_modules)
def forward(self, x: torch.Tensor) -> torch.Tensor:
B, F, T, C = x.shape
x = x[:, int(self.subband_interval[0] * F) : int(self.subband_interval[1] * F)]
x = x.permute(0, 3, 1, 2)
x = self.downsample(x)
x = self.activation(x)
x = x.permute(0, 2, 3, 1)
B, F, T, C = x.shape
x = x.reshape((B * F), T, C)
x = self.conv_modules(x)
x = x.reshape(B, F, T, C)
return x
class SDBlock(nn.Module):
def __init__(
self,
input_dim: int,
output_dim: int,
bandsplit_ratios: List[float],
downsample_strides: List[int],
n_conv_modules: List[int],
kernel_sizes: List[int] = None,
):
super().__init__()
if kernel_sizes is None:
kernel_sizes = [3, 3, 1]
assert sum(bandsplit_ratios) == 1, "The split ratios must sum up to 1."
subband_intervals = create_intervals(bandsplit_ratios)
self.sd_layers = nn.ModuleList(
SDLayer(
input_dim=input_dim,
output_dim=output_dim,
subband_interval=sbi,
downsample_stride=dss,
n_conv_modules=ncm,
kernel_sizes=kernel_sizes,
)
for sbi, dss, ncm in zip(
subband_intervals, downsample_strides, n_conv_modules
)
)
self.global_conv2d = nn.Conv2d(output_dim, output_dim, 1, 1)
def forward(self, x: torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor]:
x_skip = torch.concat([layer(x) for layer in self.sd_layers], dim=1)
x = self.global_conv2d(x_skip.permute(0, 3, 1, 2)).permute(0, 2, 3, 1)
return x, x_skip
|