Add mode method to more distribution classes #636
Conversation
@rohanbabbar04 These need to be changed if we are opting to send a tuple for multimodal purposes. |
|
@aloctavodia Issues have been updated for discrete_weibull; I will use minimize_scale from scipy.optimize since direct calculation for mode in discrete_weibull isn't feasible. For the rest, the mode is calculated as provided in Wikipedia or some reliable sources. |
| @@ -107,6 +107,14 @@ def skewness(self): | |||
| def kurtosis(self): | |||
| return num_kurtosis(self) | |||
|
|
|||
| def mode(self): | |||
There was a problem hiding this comment.
optimization routines should be inside optimization module, as they can be potentially reused. The 100 is a magic number, maybe a better option is to use self.ppf(0.9999) or similar. I tried with a couple of examples and the mode is not always right,
| @@ -1,6 +1,7 @@ | |||
| import numba as nb | |||
| import numpy as np | |||
| from scipy.stats import skew | |||
| from scipy.special import erfcinv | |||
There was a problem hiding this comment.
There is an erfcinv function implemented in the special submodule
| @@ -101,6 +102,14 @@ def entropy(self): | |||
| logpdf = self.logpdf(x_values) | |||
| return -np.trapz(np.exp(logpdf) * logpdf, x_values) | |||
|
|
|||
| def mode(self): | |||
| tau = self.nu | |||
There was a problem hiding this comment.
why renamming the variable? I get nan for some combinations of parameters like pz.ExGaussian(0, 1, 10)
| @@ -149,6 +149,37 @@ def kurtosis(self): | |||
| pdf = self.pdf(x_values) | |||
| return np.trapz(((x_values - mean) / std) ** 4 * pdf, x_values) - 3 | |||
|
|
|||
| def mode(self): | |||
There was a problem hiding this comment.
This should be in the optimization module. Do you have a reference for this method? It seems to work fine for the values I manually checked. Maybe you can check how the mode is numerically computed here https://github.com/bgctw/logitnorm
| @@ -150,6 +150,21 @@ def kurtosis(self): | |||
| * ((delta * np.sqrt(2 / np.pi)) ** 4 / (1 - 2 * (delta**2) / np.pi) ** 2) | |||
| ) | |||
|
|
|||
| def mode(self): | |||
There was a problem hiding this comment.
This gives the wrong results for large absolute values of alpha. Not sure about the proper fix, but maybe it could be enough to ensure the mode is no smaller than mu (for positive values of alpha) and no larger than mu (for negative values of alpha)
Co-authored-by: Osvaldo A Martin <[email protected]>
Description
This PR targeted #604 and work is in progress.
Implemented Modes