| 137 | |
| 138 | class SpectralConv1d(nn.Module): |
| 139 | def __init__(self, in_channels, out_channels,seq_len, modes1,compression=0,ratio=0.5,mode_type=0): |
| 140 | super(SpectralConv1d, self).__init__() |
| 141 | |
| 142 | |
| 143 | """ |
| 144 | 1D Fourier layer. It does FFT, linear transform, and Inverse FFT. |
| 145 | """ |
| 146 | |
| 147 | self.in_channels = in_channels |
| 148 | self.out_channels = out_channels |
| 149 | self.modes1 = modes1 |
| 150 | self.compression = compression |
| 151 | self.ratio = ratio |
| 152 | self.mode_type=mode_type |
| 153 | if self.mode_type ==1: |
| 154 | modes2 = modes1 |
| 155 | self.modes2 =min(modes2,seq_len//2) |
| 156 | self.index0 = list(range(0, int(ratio*min(seq_len//2, modes2)))) |
| 157 | self.index1 = list(range(len(self.index0),self.modes2)) |
| 158 | np.random.shuffle(self.index1) |
| 159 | self.index1 = self.index1[:min(seq_len//2,self.modes2)-int(ratio*min(seq_len//2, modes2))] |
| 160 | self.index = self.index0+self.index1 |
| 161 | self.index.sort() |
| 162 | elif self.mode_type > 1: |
| 163 | modes2 = modes1 |
| 164 | self.modes2 =min(modes2,seq_len//2) |
| 165 | self.index = list(range(0, seq_len//2)) |
| 166 | np.random.shuffle(self.index) |
| 167 | self.index = self.index[:self.modes2] |
| 168 | else: |
| 169 | self.modes2 =min(modes1,seq_len//2) |
| 170 | self.index = list(range(0, self.modes2)) |
| 171 | |
| 172 | self.scale = (1 / (in_channels*out_channels)) |
| 173 | self.weights1 = nn.Parameter(self.scale * torch.rand(in_channels, out_channels, len(self.index), dtype=torch.cfloat)) |
| 174 | if self.compression > 0: |
| 175 | print('compressed version') |
| 176 | self.weights0 = nn.Parameter(self.scale * torch.rand(in_channels,self.compression,dtype=torch.cfloat)) |
| 177 | self.weights1 = nn.Parameter(self.scale * torch.rand(self.compression,self.compression, len(self.index), dtype=torch.cfloat)) |
| 178 | self.weights2 = nn.Parameter(self.scale * torch.rand(self.compression,out_channels, dtype=torch.cfloat)) |
| 179 | |
| 180 | |
| 181 | def forward(self, x): |