exoplanet-dev / pymc-ext Goto Github PK

The unit vector test in exoplanet started failing (the one here will too, probably) because the default testval for that distribution gives a NaN probability. Figure out where that is going wrong and fix it!

This can use a github action then push to an orphan branch. Keep it simple!

• Optimization
• Distributions including a discussion of the math
• A more detailed comparison of tuning schedules

This depends on #1. But otherwise should be straightforward.

I am running pymc3 on a hierarchical Bayesian simulation for dust attenuation. I have a few (up to 5) independent variables and two dependent variables (which themselves are related through a prior). I have an interpolation model that compares posterior samples (data) to the model values. I have found that optimizing the likelihood and using the MAP as the starting point for NUTS sampling leads to better and quicker convergence in the chains. After I added a prior connecting the two dependent variables, I have occasionally run into the issue of the optimizer failing (presumably not coming up with anything). For example:

File "DustPymc3Interp.py", line 680, in polyNDDataBivarI
map_soln, map_info = pmx.optimize(return_info=True, start=start)
File "/mnt/home/gnagaraj/ceph/env/lib/python3.7/site-packages/pymc3_ext/optim
.py", line 135, in optimize
sys.stderr.write("message: {0}\n".format(info.message))
AttributeError: 'NoneType' object has no attribute 'message'

Here, start is simply some random configuration I assigned. Perhaps if I let return_info=False, I wouldn't get the NoneType message error, but I think the more important issue is that I am getting a NoneType in the first place. Could it just be that there is no message? That would be amazing (and annoying :-)).

I honestly don't know how to reproduce this error. It happens sporadically. For example, I have a few cases with the same properties (number of independent variables, number of galaxies and posterior samples used, etc.) but just a different set of independent variables, where one case succeeds but the other one doesn't. Here's what I would get in a successful case.

optimizing logp for variables: [rho, log_width2, log_width, ngrid, sig_n, taugrid]
message: Desired error not necessarily achieved due to precision loss.
logp: -418539.35894535476 -> 83080.45132546293
optimizing logp for variables: [rho, log_width2, log_width, ngrid, sig_n, taugrid]

• Version of PyMC3: 3.11.2
• Version of PyMC3-ext: 0.1.0
• Operating system: GCC 4.8.5 20150623 (Red Hat 4.8.5-44)] on linux
• Python version & installation method (pip, conda, etc.): 3.7.3 (through pip)

Please let me know if you would like any other information. I mainly just wanted to let you know that it is possible for the optimizer to fail in weird ways.

A lot of these changes, specifically those for tuning would be a great addition to PyMC3 directly. Have you considered PRing them upstream?