About Anormal detection

[nixwl]

at /ModernTCN/[ModernTCN-detection/models/ModernTCN.py , function forward_feature -> Source Here

 def forward_feature(self, x, te=None):

        B,M,L=x.shape
        x = x.unsqueeze(-2)

        for i in range(self.num_stage):
            B, M, D, N = x.shape
            x = x.reshape(B * M, D, N)

            if i==0:
                if self.patch_size != self.patch_stride:
                    # stem layer padding
                    pad_len = self.patch_size - self.patch_stride
                    pad = x[:,:,-1:].repeat(1,1,pad_len)
                    x = torch.cat([x,pad],dim=-1)
                x = x.reshape(B,M,1,-1).squeeze(-2)
                x = x.unfold(dimension=-1, size=self.patch_size, step=self.patch_stride)
                x = self.downsample_layers[i](x)
                x = x.permute(0,1,3,2)


            else:
                if N % self.downsample_ratio != 0:
                    pad_len = self.downsample_ratio - (N % self.downsample_ratio)
                    x = torch.cat([x, x[:, :, -pad_len:]],dim=-1)
                    x = self.downsample_layers[i](x)
                    _, D_, N_ = x.shape
                    x = x.reshape(B, M, D_, N_)

            x = self.stages[i](x)
        return x

It seems that multi-scale downsampling is not implemented here? Because if N % self.downsample_ratio == 0, then x = self.downsample_layers[i](x) will not be executed. Is this an indentation error? The condition if N % self.downsample_ratio != 0: only performs padding on x.
The modified code might be as follows:

 else:
    if N % self.downsample_ratio != 0:
        pad_len = self.downsample_ratio - (N % self.downsample_ratio)
        x = torch.cat([x, x[:, :, -pad_len:]],dim=-1)
    x = self.downsample_layers[i](x)
    _, D_, N_ = x.shape
    x = x.reshape(B, M, D_, N_)