New pymc3 user here :slight_smile: I’ve been trying to get a slightly modified version of this pymc3 GLM logistic regression tutorial to work - to no avail. print(pm.__version__) 3.6 I want to use custom (non-default) priors for the GLM coefficients. Here’s my code: with pm.Model() as logistic_m…

It looks more like a numpy version problem. Try upgrading numpy and scipy, and then try again.

I’m running this on Google’s colab platform: print(np.__version__) 1.14.6 import scipy print(scipy.__version__) 1.1.0 Updated both numpy and scipy to the latest versions 1.16.1 and 1.2.0 (with pip). I get the same error.

Thank you! This got me a bit further: with pm.Model() as logistic_model: my_priors = {"Intercept": pm.Normal.dist(mu=0, tau=1e-2), "age": pm.Uniform.dist(0,100), "age2": pm.Uniform.dist(0,10000), # statistically not correct, but whatever "educ": pm.…

The uniform prior for age2 is way too wide… but the Bad initial energy might due to the start point jitter in the default initialization. Try: trace_logistic_model = pm.sample(2000, chains=4, tune=1000, init='adapt_diag')

Ran with "age2": pm.Uniform.dist(0,3500) (age2 := age^2) and trace_logistic_model = pm.sample(2000, chains=4, tune=1000, init='adapt_diag') Same error…

In that case, check the input table (make sure there is no NaN or Inf), then check the output of logistic_model.check_test_point() make sure there is no NaN

I think we may be onto something… The check assert data.isnull().values.any() == False passes. (data[['age','educ','hours','income']] >= 0).all(1).all() so there are no negative values in those data columns. However, here is what logistic_model.check_test_point() returns: Intercept 0.0…

Most likely after multiplication and transformation the linear predictor is at the boundary (i.e., p becomes 0 or 1), try rescale your predictors.

All right! I scaled the predictors. Now the NUTS sampler seems to work. I’ll have a look at the traces and then report back. Many thanks!

GLM logistic regression with custom prior in pymc3 (v. 3.6)

Questions

junpenglao February 1, 2019, 2:53pm 4

Also your way of defining the prior is not correct, see: Linear Regression With Positivity Constraint

Topic		Replies	Views
Using prior distributions in pymc3 Questions development	0	677	May 30, 2019
Some fun with GitHub Copilot and PyMC Sharing	2	653	March 23, 2022
Raise ValueError('Cannot use scalars') Questions	12	2021	December 12, 2017
What exactly are the default prior specification for glm.from_formula() patsy style? Questions	1	946	May 6, 2020
Simple logistic model: Bad initial energy log p's -Inf and Inf Questions	2	430	July 28, 2021

GLM logistic regression with custom prior in pymc3 (v. 3.6)

Related topics