human-analysis / intrinsic-dimensionality Goto Github PK

Hi,

I'm trying to use DeepMDS on VOC for a self supervised learning task, however it's going down by a lot from the baseline when we even try to do 9600 -- 8192, see 50% drop, I'm not sure why this would happen. As you guys proved, there was nearly all information retention at intrinsic space, so I'm a bit confused.

There is no mention in the paper about the genuine and imposter feature representations.
Also, the algorithm of DeepMDS is supposedly unsupervised so what are the labels being used for?

Hi:
in your paper, it should be $y=log(\frac{p(r)}{p(r_{max})})$, but in your code, in my opinion, (4ac-b**2)/4./a is to fit $p(r_{max})$, so the code should be :

left_distr_y= np.log(distr_y[np.logical_and(np.logical_and(distr_x[:]>rmin,distr_x[:]<=rM),distr_y[:]>0.000001)])-np.log((4*a*c-b**2)/4./a)

Or just I make a confuse?

And fuc2 is to fit which function? (5)?

    fit = curve_fit(func2,left_distr_x,left_distr_y)
    ratio=np.sqrt(fit[0][0])

And the ratio is to refer d? or is there anything I make a mistake?

Thank you!
Best Wishes!

Hello, in your func, b and c doesn't mean anything, but in #L286

y2=func(left_distr_x,fit[0][0],fit[0][1],fit[0][2])

I think is there some missing?

Best wishes,

How do you get the feature of LFW? In your table 2, 512-dim only achieve 96.74%, but when I look at the implement of sphereface, it could achieve 99.30%. And when I test the LFW feature the sphereface released, it could achieve 98.02%, and, when I use PCA to reduce the dimension to 16, it still achieve 0.8481, which is higher than your reported 32.67%.

So I have a question: where does your feature come from? Could you give me your lfw feature? I want to reproduce the table 2.

Hello,
when I run ID_graph_largedata.py, the console reported,
'''
FileNotFoundError: [Errno 2] No such file or directory: '/research/prip-gongsixu/results/feats/evaluation/feat_lfwblufr_sphere.mat',
'''
I think the mat document is needed, or how to solve the problem?

Hi:
I have another problem about equation 4 and code.
$$d_H(x_i,x_j)-d_L(y_i, y_j)$$
in my opinion, $d_H$ is to use geodesic distance, and $d_L$ is to use Euclidean or cosine distance.

but in your code,

        if y1 is not None:
            if self.dist_metric == 'cosine':
                dist_in = torch.sum(torch.mul(x1,x2) / (torch.norm(x1,dim=1,keepdim=True)
                    *torch.norm(x2,dim=1,keepdim=True)), dim=1)
                dist_out = torch.sum(torch.mul(y1,y2) / (torch.norm(y1,dim=1,keepdim=True)
                    *torch.norm(y2,dim=1,keepdim=True)), dim=1)
            elif self.dist_metric == 'Euclidean':
                dist_in = torch.norm(x1-x2, dim=1, keepdim=True)
                dist_out = torch.norm(y1-y2, dim=1, keepdim=True)
            else:
                raise(RuntimeError('Metric does not support!'))

So, in your paper, it is that $d_H$ and $d_L$ both use Euclidean, so is there any problem?

Line 233 in ID_graph_largedata.py
std=np.sqrt(abs(-1/a0/2.))
It is easy to understand the rmax can be via -b/a/2 according to the property of quadratic equation of one unknown, but it is hard to understand the standard deviation is estimated in this way?
Thanks.