def forward()

in Dassl.pytorch/dassl/modeling/ops/efdmix.py [0:0]

• 24 lines of code
• 7 McCabe index (conditional complexity)

    def forward(self, x):
        if not self.training or not self._activated:
            return x

        if random.random() > self.p:
            return x

        B, C, W, H = x.size(0), x.size(1), x.size(2), x.size(3)
        x_view = x.view(B, C, -1)
        value_x, index_x = torch.sort(x_view)  # sort inputs
        lmda = self.beta.sample((B, 1, 1))
        lmda = lmda.to(x.device)

        if self.mix == "random":
            # random shuffle
            perm = torch.randperm(B)

        elif self.mix == "crossdomain":
            # split into two halves and swap the order
            perm = torch.arange(B - 1, -1, -1)  # inverse index
            perm_b, perm_a = perm.chunk(2)
            perm_b = perm_b[torch.randperm(perm_b.shape[0])]
            perm_a = perm_a[torch.randperm(perm_a.shape[0])]
            perm = torch.cat([perm_b, perm_a], 0)

        else:
            raise NotImplementedError

        inverse_index = index_x.argsort(-1)
        x_view_copy = value_x[perm].gather(-1, inverse_index)
        new_x = x_view + (x_view_copy - x_view.detach()) * (1-lmda)
        return new_x.view(B, C, W, H)