| 121 | |
| 122 | |
| 123 | class UpConv(nn.Module): |
| 124 | def __init__( |
| 125 | self, |
| 126 | embed_dim: int = 768, |
| 127 | stride: int = 4, |
| 128 | ): |
| 129 | super().__init__() |
| 130 | self.embed_dim = embed_dim |
| 131 | self.stride = stride |
| 132 | self.in_proj = nn.Linear(embed_dim, self.stride * embed_dim) |
| 133 | # Simple transpose convolution layer to keep channel number consistent |
| 134 | self.up_conv = nn.ConvTranspose1d( |
| 135 | self.stride * embed_dim, |
| 136 | embed_dim, |
| 137 | kernel_size=stride, |
| 138 | stride=stride, |
| 139 | bias=False, |
| 140 | ) |
| 141 | |
| 142 | def forward(self, x: torch.Tensor, input_length: torch.Tensor): |
| 143 | x = self.in_proj(x) |
| 144 | x = x.transpose(1, 2) |
| 145 | res = self.up_conv(x) |
| 146 | res = res.transpose(1, 2) |
| 147 | output_length = input_length * self.stride |
| 148 | return res, output_length |
| 149 | |
| 150 | |
| 151 | class RedCodec(nn.Module): |