PyMC3.2 on bleeding edge theano has significantly slower compilation time


#1

Did not try to replicate but here are some details on what happened, not a large issue but could perhaps be part of something bigger with theano backend

My Specs: Windows 10 with Jupyter notebook running Pymc.

Before: Fresh install of anaconda + PyMC3.2 from Conda forge. compilation time was reasonable.
After:

  1. Was getting dynamic library warnings from BLAS (as well as blas flag errors) when using tt.dot
  2. Scared that all my current computations are numpy based, I decided to update theano to bleeding edge (suggested by many others) and install pymc3.2 (release vers) through pip.
  3. Compilation time for the exact same models are now significantly longer (about 5-10x longer)

#2

Thanks for reporting - anybody else has the same issue?


#3

I didn’t investigate, but the last couple of days I was pretty surprised that theano seemed to take ages for a couple of things. Not sure it was 5-10x however, and I didn’t compare directly.


#4

I just upgraded to 3.2 and I am experiencing a similar slowdown. Should we post an issue?


#5

Did you also upgrade to theano 1.0?
I opened an issue here: https://github.com/pymc-devs/pymc3/issues/2723
If you recently also upgrade theano please post your theano version to there :wink:


#6

Yest, I upgraded to theano 1.0, sorry I forgot to mention that.