scikit-stan

What is scikit-stan

scikit-stan will enable you to use various bayesian models based on stan(http://mc-stan.org) and pystan with an elegant interface like a scikit-learn or keras.

Demo

import numpy as np

from skstan.regression.linear_models import LogisticRegression

if __name__ == '__main__':
    x = np.array(
        [
            [1,2,3,],
            [1,2,7,],
            [1,0,3,],
            [1,1,3,],
            [3,7,3,],
        ]
    )
    y = np.array([0,0,0,0,1])

    glm = LogisticRegression(shrinkage=10, chains=8)
    fit = glm.fit(x, y)

Then we got result object fit, and field stanfit is a stanfit object of pystan.

print(fit.stanfit)

It gives following

Inference for Stan model: anon_model_f63cd5ccdd67c22034b2490ae4c9cdd1.
4 chains, each with iter=2000; warmup=1000; thin=1; 
post-warmup draws per chain=1000, total post-warmup draws=4000.

           mean se_mean     sd   2.5%    25%    50%    75%  97.5%  n_eff   Rhat
alpha[0]   2.23    0.29   8.88 -14.39  -3.87   2.02   8.07  20.67    966    1.0
alpha[1]   7.81    0.18   5.29  -1.08   4.01   7.48  11.23  19.01    880    1.0
alpha[2]  -9.79    0.22   5.87 -22.91 -13.41  -9.37  -5.38  -0.17    728    1.0
beta      -2.48    0.29    9.8 -22.63  -9.03   -2.3   3.99  16.91   1146    1.0
yp[0]    -13.99    0.32  11.19 -40.69 -20.24 -11.35  -5.42    0.3   1259    1.0
yp[1]    -53.15    1.14  32.08 -128.4 -71.99 -48.46 -29.35  -5.24    790    1.0
yp[2]    -29.61     0.6  16.66 -67.24 -39.44 -27.97  -17.0  -4.37    771    1.0
yp[3]     -21.8    0.44  13.17 -52.98 -29.03 -19.57 -11.91  -3.23    894    1.0
yp[4]     29.51    0.69  24.68   0.58   10.3  23.36  42.72  90.17   1276    1.0
lp__      -2.16    0.05   1.48  -5.93   -2.9  -1.81  -1.07  -0.32    956    1.0

Samples were drawn using NUTS at Thu Apr 13 07:52:33 2017.
For each parameter, n_eff is a crude measure of effective sample size,
and Rhat is the potential scale reduction factor on split chains (at 
convergence, Rhat=1).

Result object of skstan also have prediction methods. Predicted values can be obtained as samples from distribution with a predict_dist method, because it is bayesian model.

yp_dist = fit.predict_dist(x)
print(yp_dist)

Then we got

array([[  2.63886682e-08,   5.23976746e-04,   5.54863097e-05, ...,
          2.46008578e-08,   3.74830192e-01,   3.45994043e-03],
       [  1.07746578e-22,   1.01664809e-18,   4.12813154e-26, ...,
          5.64992544e-19,   7.24386097e-12,   1.75795155e-23],
       [  8.04688037e-22,   4.44522113e-12,   1.42920488e-11, ...,
          7.71565191e-13,   5.13118658e-05,   4.26331280e-05],
       [  4.60810657e-15,   4.82743551e-08,   2.81612678e-08, ...,
          1.37772153e-10,   5.51614998e-03,   3.84594197e-04],
       [  9.99999998e-01,   1.00000000e+00,   1.00000000e+00, ...,
          9.99965378e-01,   1.00000000e+00,   1.00000000e+00]])

So let's check the histgram of first row with pandas.Series.

import pandas as pd
pd.Series(yp_dist[0]).hist(bins=20)

If you need a median of samples, you can get it with just predict method

yp = fit.predict(x)
print(yp)

gives

array([  1.17280235e-05,   9.01419773e-22,   7.16023732e-13,
         3.18368664e-09,   1.00000000e+00])

How to install

Install

Installers for the latest released version are available at PyPI.

pip3 install skstan

Install from sources

git clone https://github.com/BayesianFreaks/scikit-stan
cd scikit-stan
python3 setup.py install

Uninstall

pip3 uninstall scikit-stan

Using python2?

Are you joking?

We can't touch you because we are living in the future from you, and you're living in past ages. Please say hello to Nobunaga Oda.

We will always use newest features of the latest version of python, so you should use the latest version of python.

Models

Ready

Regression Models

Linear Regrassion
Poisson Regression
Logistic Regression

Next Steps

Regression Models

Gamma Regression
GLMM
etc...

Time Series Models

AR Model
MA Model
ARMA Model
ARIMA Model
ARCH Model
GARCH Model
TAR Model
State Space Model
or Some Dynamic Regression Models
etc...

Clustering Model

Gaussian Mixture Model
Latent Dirichlet Allocation
etc...

Particular Application

Modeling about online-advertisement
Decompose time series data
Empirical Bayesian Estimation