Why scale the outputs by 30 in Decoder.update_buffer?

[arvieFrydenlund]

The 30 here seems to be a magic number unless I missed something in the paper?

    def update_buffer(self, S_tm1, c_t, o_tm1, ident):
        # concat previous output & context
        idt = torch.tanh(self.F_u(ident))
        o_tm1 = o_tm1.squeeze(0)
        z_t = torch.cat([c_t + idt, o_tm1/30], 1)
        z_t = z_t.unsqueeze(2)
        Sp = torch.cat([z_t, S_tm1[:, :, :-1]], 2)

        # update S
        u = self.N_u(Sp.view(Sp.size(0), -1))
        u[:, :idt.size(1)] = u[:, :idt.size(1)] + idt
        u = u.unsqueeze(2)
        S = torch.cat([u, S_tm1[:, :, :-1]], 2)

        return S

Thanks.

[macarbonneau]

I would like to express my interest regarding this question.
Also, while on the subject of the memory buffer update, I would like to know where is implemented the teacher forcing idea described in section 3.1 of the paper. I do not see where the predicted output is mixed with the noisy target.