This package is dedicated to realizing methods used in this paper.
A package that transforms anything to a Gaussian process.
pip install CWGP
Let's randomly generate 100 numbers following an exponential distribution.
import numpy as np
import seaborn as sns
import matplotlib.pyplot as plt
exp = np.random.exponential(scale=5, size=50)
idx = np.arange(50)
We now instantiate a CWGP class consisting of 3 Sinh-Arcsinh transformations.
from cwgp.cwgp import CWGP
compgp = CWGP(["sa","box_cox"])We then fit our data into the model. This minimizes the negative log likelihood function and stores the corresponding parameters for us.
compgp.fit(exp, idx)To get the parameters, we do
params = compgp.phi.res.xWe then transform the data via
t_exp, d = compgp.phi.comp_phi(params, exp)
sns.distplot(t_exp)
plt.show()
Let's make a QQ-plot and see how Gaussian it is.
from scipy import stats
stats.probplot(t_exp, dist="norm", plot=plt)
plt.show()
The inverse function is also implemented.
inv_t_exp = compgp.phi.inv_comp_phi(params, t_exp)fig, ax = plt.subplots(1, 2)
sns.distplot(inv_t_exp, ax=ax[0])
sns.distplot(exp, ax=ax[1])
plt.show()The one on the left is the one being transformed and transformed-back, and the one on the right is the original distribution.
They should be exactly the same.
from cwgp.transformations import sa
from cwgp.transformations import asinh
from cwgp.transformations import box_cox
from cwgp.transformations import sal