rte-france / relife Goto Github PK

Should allow to customized label names when plotting ECDF:

# Import load_circuit_breaker function from relife.dataset
from relife import *
from relife.datasets import load_circuit_breaker 

# Assign time, event variables from dataset
time, event = load_circuit_breaker().astuple()[:-1]

# Get only time values for observed event
time1 = time[event==1]

# Fit ECDF estimator to time1
ecdf = ECDF().fit(time1)

# Plot ECDF estimator of F and 1-F
ecdf.plot(label='custom_label')

returns error:
TypeError: plot() got multiple values for keyword argument 'label'

The computation of the derivative of the upper gamma function in

def _jac_uppergamma_c(c: np.ndarray, x: np.ndarray) -> np.ndarray:
     return shifted_laguerre(lambda s: np.log(s) * s ** (c - 1), x, ndim=np.ndim(x))

is unstable for low values of the parameters. The numerical result diverges from the true value as parameters go to 0. Test example : c = 0.001, x = 0.002.

returns an error printing variance for a Gamma distribution.
test:

from relife import *
gamma = Gamma(c = 41.26, rate = 1.62)
gamma.var()

returns:

TypeError: var() missing 1 required positional argument: 'params'

should returns the variance of the gamma distribution as a float.

temporary solution:

gamma.var(gamma.params)

A proposition to fix this issue:
In distribution.py:

    def var(self) -> np.ndarray:
        c, rate = self.params
        return c / (rate**2)

"Fit the empirical cumuative distribution function."
instead of:
"Fit the empirical cumulative distribution function."

Some issues may happen when fitting data. There is a convergence problem using with the minimize function of scipy.
Here is an example with Gompertz distribution:

from relife import *

weibull = Weibull(c=7.531, rate=0.037)
rtf = RunToFailure(weibull, cf=1)

datagen = rtf.sample(T=300, n_samples=100, random_state=0)
time, event, entry = datagen.to_lifetime_data(t0=0,tf=30).astuple()
gompertz = Gompertz().fit(time,event,entry)

/home/xxx/.local/lib/python3.8/site-packages/relife/parametric.py:347: RuntimeWarning: invalid value encountered in sqrt
self.se = np.sqrt(np.diag(self.var))

Problem may be solved by improving the initialization "params0" and/or using another type of solver (by default L-BFGS).

When plotting a ParametricLifetimeDistribution, the plot label is by default the name of the distribution.
Would be better to add the name of the function (=fname) by default (ex: "Exponential sf", instead of "Exponential").

is : f(t) = c \lambda^c t^{c-1} e^{-(\lambda t)^c}
should be : f(t) = c \lambda (\lambda t)^{c-1} e^{-(\lambda t)^c}
(right formula for a parametrization with rate parameter \lambda).

for ar in OneCycleAgeReplacementPolicy and ar1 for AgeReplacementPolicy:
should be "times to preventive replacements" or "times until preventive replacements" (or sthg else) instead of "Ages of the first preventive replacements"

Hello,

I am fitting a Kaplan-Meier estimator with doubly censored data. But I get an error. To reproduce the error use the example below.

from relife.nonparametric import KaplanMeier as KM
km = KM()
km.fit(time=np.array([1,2,3]),event=np.array([1,0,2]))

The error is:

  File ~/relife/nonparametric.py:54 in _estimate
    raise ValueError("event values must be in [0,1]")
ValueError: event values must be in [0,1]

But the documentation says that event 2 is used to flag an event as left-censored.