About supporting safe_sign_sqrt

[nguyenthikimtruc]

Thank you for support the smsop gitbub.
Could you please sunport the function safe_sign_sqrt for normalization step of class SignedSqrt(Layer).
and in elif output_normalization == 'cubicroot':
x = PowLayer(scale=0.5, center=np.power(49. / 2 - 1.0 / 3, 1.0 / 3))(x) , what is the meaning of the number 49. / 2 - 1.0 / 3
Thank you,
Truc

[kcyu2014]

Sorry for the late reply. These numbers comes from Wilson–Hilferty transformation. In short, doing a cubic root with this mean and std will make the previous Chi distribution converge to Gaussian.

For details, please refer to

Johnson, N. L.; Kotz, S.; Balakrishnan, N. (1994). "Chi-Square Distributions including Chi and Rayleigh". Continuous Univariate Distributions. 1 (Second ed.). John Wiley and Sons. pp. 415–493. ISBN 978-0-471-58495-7.