| 53 | return self.pe[:, scale:x.size(1)*scale+1:scale] |
| 54 | |
| 55 | class TokenEmbedding(nn.Module): |
| 56 | def __init__(self, c_in, d_model): |
| 57 | super(TokenEmbedding, self).__init__() |
| 58 | padding = 1 if torch.__version__ >= '1.5.0' else 2 |
| 59 | self.tokenConv = nn.Conv1d(in_channels=c_in, out_channels=d_model, |
| 60 | kernel_size=3, padding=padding, padding_mode='circular', bias=False) |
| 61 | for m in self.modules(): |
| 62 | if isinstance(m, nn.Conv1d): |
| 63 | nn.init.kaiming_normal_(m.weight, mode='fan_in', nonlinearity='leaky_relu') |
| 64 | |
| 65 | def forward(self, x): |
| 66 | x = self.tokenConv(x.permute(0, 2, 1)).transpose(1, 2) |
| 67 | return x |
| 68 | |
| 69 | |
| 70 | class FixedEmbedding(nn.Module): |