Cyclic FIN Layer to Linear FIN Layer

[cyiheng]

Hi !

First of all thank you for your work, I was waiting your code since I read your paper !

I was wondering if you can give some advice to modify your code from a cyclic function of the FIN layer to a Linear one ?

I actually try to only replace every single cos_phi / sin_phi to a simple phi, but I'm not sure that will be enough. Maybe I will miss some major points by only changing these.

Thank you again !

[fabvio]

Hi! Thank you for your interest. First of all, this is a simple implementation of linear FIN that you can use as reference.

class LinearFIN(nn.Module):
    def __init__(self, dim):
        super().__init__()
        self.instance_norm = nn.InstanceNorm2d(dim, affine=False)
        self.a_gamma = nn.Parameter(torch.zeros(dim))
        self.b_gamma = nn.Parameter(torch.ones(dim))
        self.a_beta = nn.Parameter(torch.zeros(dim))
        self.b_beta = nn.Parameter(torch.zeros(dim))

    def forward(self, x, phi):
        gamma = self.a_gamma.unsqueeze(0) * phi.unsqueeze(-1) + self.b_gamma.unsqueeze(0)
        beta = self.a_beta.unsqueeze(0) * phi.unsqueeze(-1) + self.b_beta.unsqueeze(0)
        return self.instance_norm(x) * gamma.unsqueeze(-1).unsqueeze(-1) + beta.unsqueeze(-1).unsqueeze(-1)

As regards the modifications to the rest of the code, you should be mostly fine if you just replace the cos_phi and sin_phi with a single phi. The only point where you should be more cautious is in the phiNet_A angle estimation, which should be replaced by a point estimation on a linear manifold. Hence, you should avoid extracting sine and cosine of the estimated angle and you should use directly the predicted value.