khammernik / sigmanet Goto Github PK

Hello Kerstin,

It is really a great job~
I want to reproduce your work, but I am wandering if one or two 1080Ti can handle with the huge GPU memory consumption, as mentioned in the paper.

So could you show me your environment? Thanks a lot.

Best Regards,
He Jin Qiang

Dear @khammernik ,

I've got a question/comment concerning back propagation of the proximal gradient layer with respect to lambda.
I got curious, reading your current MRM paper where you wrote that training becomes unstable when lambda is not fixed.

Following your conventions,
M := lambda A^HA + 1
Q := M^-1.

Now, the the derivative of the inverse of a matrix (with respect to lambda) is given by:
Q'=-Q M' Q

In the code, I see twice the Q as expected but not M' = A^HA.
Is it missing or does it cancel somehow?

Best regards,
Moritz

Hello @khammernik @js3611 @j-duan @cq615 ,

Excellent result is achived from SN, however, there is still some problem in PCN.
I notice that the following class has not been completed, however, which is necessary for PCN.
Would you be so kind to help me?

class ParallelComplexInstanceNorm(nn.Module):
    """ Complex normalise each parallel image """
    def __init__(self, num_chans):
        super(ParallelComplexInstanceNorm, self).__init__()
        self.num_chans = num_chans
        self.means = torch.from_numpy(np.repeat([0], self.num_chans))
        self.covs = torch.from_numpy(np.repeat(
            [1 / np.sqrt(2), 0, 0, 1 / np.sqrt(2)],
            self.num_chans,
        ))

    def matrix_invert(self, xx, xy, yx, yy):
        pass

    def complex_instance_norm(self, x, eps=1e-5):
        pass

    def complex_pseudocovariance(self, data):
        pass

    def forward(self, input):
        pass

    def set_normalization(self, mean, cov):
        pass

    def normalize(self, x):
        return x

    def unnormalize(self, x):
        return x

thx a lot,
kingaza

Hello @khammernik ,

Would you mind if sharing the trained model?
Grateful for you help!

Best Regards,
kingaza