bug in ODEFnc()

[jiangxinke]

There is a bug in ODEFnc(), for the Parameter can not be augmented assignment directly, or else it would have an error "cannot assign ‘torch.cuda.FloatTensor’ as parameter ‘self.w’ (torch.nn.Parameter or None expected)"

it should be :

        self.w.data = (1 + self.beta) * self.w - self.beta * torch.mm(torch.mm(self.w, torch.t(self.w)), self.w)
        xw = torch.einsum('ijkl, lm->ijkm', x, w)

        d2 = torch.clamp(self.d2, min=0, max=1)
        w2 = torch.mm(self.w2 * d2, torch.t(self.w2))
        self.w2.data = (1 + self.beta) * self.w2 - self.beta * torch.mm(torch.mm(self.w2, torch.t(self.w2)), self.w2)

[square-coder]

You are right, but we found such operation led to suboptimal convergence. And a simple alternative is to just delete this line, which means we don’t constrain $self.w$ to be orthogonal, and it works in practice. We have fixed the bug in this way.
Thanks for your attention.

[jiangxinke]

tks