Open jackyyigao opened 1 month ago
Hi, I think the codes below actually play the role of power law compressing.
x_mag = torch.sqrt(x[...,[0]]**2 + x[...,[1]]**2 + 1e-12)
x_c = torch.div(x, x_mag.pow(1-self.c) + 1e-12)
Actually, the codes are equal to such a version, which may be more intuitional.
x_mag = torch.sqrt(x[...,[0]]**2 + x[...,[1]]**2 + 1e-12)
x_c_mag = x_mag.pow(self.c)
x_c_real = x_c_mag * x[...,[0]] / (x_mag + 1e-12)
x_c_imag = x_c_mag * x[...,[1]] / (x_mag + 1e-12)
x_c = torch.cat([x_c_real, x_c_imag], dim=-1)
I hope I have clarified your confuse, If not, feel free to ask me.
thank you very much for your so quick response! BTW, your latest work on GTCRN seems very promising for real-time implementation for commercial products. hope could will have change to communicate with you later.
Thank you for your appreciation, and I'm glad to keep in touch with you
Hi, Xiaobin, the FE block in you implementation seems a kind of normalization to the complex spectra, could you help introduce the paper to describle it? Seems it's not the power law compressed complex spectra as mentioned in the deepVQE paper? thanks for any info.