About re-scoring analysis

[genforce]

About re-scoring analysis

[ShenYujun]

You can set k=1 for all directions for a fair comparison.

[rainsoulsrx]

hi，In this table, you mentioned " each level of layers", for example, Middle Layers (2-5), means 2-5 layers together as a whole matrix or a matrix for each layer? I think it must satisfy linearity, but 2-5 layers together is not linearity because there is nonlinear activation function between them.

[ShenYujun]

If you only consider how the latent code is fed into these AdaIN operations for each layer, they are linear.

[07hyx06]

Suppose layer2's weight matrix size is [1024,512], layer3's size is [512,512], layer 4's size is [256,512]. If consider layer2~4, it firstly concatenate them together and get a [1024+512+256,512]=[1792,512] size matrix A, and compute A^T A's eigen vector as the direction?

[ShenYujun]

@07hyx06 Yes, you are right.

[07hyx06]

@07hyx06 Yes, you are right.

I concatenate all of the 18 layers together and get a [10176,512] weight matrix A, and then normalized A's row by A[i]/=torch,norm(A[i]) before performing eigen decomposition, but i still cannot get the results in Table 1. The cos distance in my implement is 0.8+ for all attribute. Is it needs other tricks? I used torch.symeig to compute the eigens and the biggest eigen is 60 6.55.

[ShenYujun]

Please try np.linalg.eig(A.T.dot(A))