alelab-upenn / graph-scattering-transforms Goto Github PK

In the definition of the Hann kernel, 'm a' (as appears in eq(4) from Shuman et. al. and defined in eq. (10)) is defined as

Later on, for the case without warping you define

This doesn't seem right. Let's say J = 3, then t has size (3,). Then you call t[3], which is out of bound for t with size (3,).

Hello, in the implementation of the Hann wavelet, the summation term at the end of the warped scaling function (eq. 13 from Shuman et. al.) seems to be calculated for the last scale only. I think you should change

to

sumPsiSquared += np.abs(psi) ** 2

Hello, I was wondering why the low-pass filter U in the diffusion wavelet is defined as a vector of shape (N), N being the number of nodes:

self.U = d/np.linalg.norm(d, 1)

I was expecting U to be a square matrix. I found that Zou and Lerman (which you have cited) are also defining U as a square matrix of size (N, N).
U = np.linalg.matrix_power(T, t)

The fact that you are using a vector also gives a different number of output features.

Same question goes for tight Hann wavelets where low-pass filter U is

self.U = (1/self.N) * np.ones(self.N)

In scattering transform, scaling function is used as low pass filter during scattering transform to capture low frequency or average behavior of the signal. I was expecting that you use equation (13) of Shuman et. al. as low-pass filter, and use equation (12) as wavelets, but it seems that you have used both relations as wavelets and used self.U = (1/self.N) * np.ones(self.N) as low-pass filter.

1. Isn't scaling supposed to be the low pass filter that computes invariants and wavelets the high pass filters?

2. using a vector as opposed to a squared matrix makes your implementation of Diffusive wavelets have different dimensionality of output features in comparison with Zou and Lerman. Which one should I trust?