ImportError: cannot import name 'FunctionReduce'

[jcr-lyxor]

Hi

I am trying to use the package, I have installed using conda and get the following error:

from optimix import FunctionReduce ImportError: cannot import name 'FunctionReduce

Although I understand how EP works I am still struggling in the implementation. Would you have a simple example using Logistic or Poisson regression with Gaussian priors please?

[horta]

Hi @jcr-lyxor ,

Does it solve if you do conda update glimix-core optimix ?

Have you seen the documentation?

[jcr-lyxor]

No it doesn't. FunctionReduce does not seem to be a part of optimix .

I had a look in the documentation. I just wonder if you could provide such simple example to help understand how it works.

[horta]

what are the versions of the installed glimix-core and optimix packages?

glimix_core.__version__ and optimix.__version__

[jcr-lyxor]

respectively 1.6.0and 3.0.3 under python 3.5.6.

[horta]

The current version of glimix-core is 3.1.7. It supports python 3.6 onwards but you could try it on 3.5 and see if it works: conda update -c conda-forge glimix-core

[jcr-lyxor]

It does't install/update 3.1.7, it is still glimix-core-1.6.0. I am under windows 10.

[horta]

glimix-core is platform independent. It might be because of python 3.5. Try: conda install -c conda-forge "glimix-core==3.1.7" What does it show?

Also, you can install glimix-core via pip. It works the same.

[jcr-lyxor]

I have issues with pip as well. For some reason liknorm is falling at install ERROR: Command "python setup.py egg_info" failed with error code 1

[horta]

I'm skipping py 3.5 build for liknorm python package: https://github.com/limix/liknorm-py/blob/master/.travis.yml#L6 I suggest updating your python: removing conda and installing it again from https://www.anaconda.com/distribution/ It will come with Python 3.7.

[jcr-lyxor]

Ok I try with 3.7 env. Thanks for looking a it.

About EP with for GLM with Gaussian prior would you have a simple example or/and could I solve it with this lib?

[horta]

See if https://glimix-core.readthedocs.io/en/latest/glmm.html helps.

[jcr-lyxor]

All good under py37. Thanks a lot.

I guess with G=0 I have what I want. Still if you have a simple script that helps to understand the steps that would be appreciated.

I

[jcr-lyxor]

Actually I am not sure because you are using Gaussian Processes under the wood for the moment matching?

For instance the following simulation gives me a solution far away:

from glimix_core.glmm import GLMMExpFam
import numpy as np

np.random.seed(115283)
beta_0 = 0.7
beta_1 = -1 
beta_2 = 0.5

N = 500
x = np.random.normal(0, scale=1, size=N)
x2 = np.random.normal(0, scale=1, size=N)

X = np.array([[1]*N, x, x2]).T

beta = np.array([beta_0, beta_1, beta_2])
true_mu = np.exp(X@beta)
n = 1
p = true_mu / (1 + true_mu)
y = np.random.binomial(n=n, p=p, size=N)

glm = GLMMExpFam(y,('Bernoulli' ),X)
glm.fit(verbose=False)
print(glm.beta)
[ 10.62991091 -15.15332283   7.80204789]

while statsmodel gives the right answer:

import statsmodels.api as sm
glm = sm.GLM(y,X,family=sm.families.Binomial())
res = glm.fit()
res.params
[ 0.88751168, -1.27477813,  0.65629759]

Am I missing something?

[horta]

Please, show the output that those lines produce.

Also, pay attention that GLMMExpFam is modelling a more complex model: there is an additional latent iid noise. (i dont know how sm.GLM is, but doesnt seem to include it). Make use of glm.v0 and glm.v1 to understand more the model and have a look at the third equation of https://glimix-core.readthedocs.io/en/latest/glmm.html .

[horta]

Also, note that you can set and keep fixed the variances of GLMM model.

[jcr-lyxor]

I have updated. Indeed GLMMExpFam is far more complex that is was I was asking for a simpler version of GLM+Gaussian prior before using it. I guess that corresponds to glm.v0 = 0 and glm.v1=1 and I do not want the latent so G=0. Stats model is just the normal maximum likelihood logistic regression.