evidation-health / ContinuousTimeMarkovModel

Continuous-time Markov model with discrete observations
MIT License
23 stars 23 forks source link

Out of bounds error when launching parallel jobs #13

Open pourzanj opened 9 years ago

pourzanj commented 9 years ago

Why is there an out of bounds error in Theano when running njobs > 1

pourzanj commented 9 years ago

Running with njobs > 1 will create Theano compile errors that are not there when njobs=1. Numpy out of bounds errors can also be introduced when njobs > 1. The Theano errors are mostly related to type mismatches such as an int64 when there should be an int32.

When all the sections of the code that cause such errors are commented out, the program runs but the time to draw a simple sample is much slower. Originally samples took 2 seconds but with njobs >1 they take 7, even though they run in parallel. A Pymc3 issue was opened here.

dstuck commented 9 years ago

When fastcompile is turned on, sometimes new int typing errors occur in theano. Maybe this a reordering of operations or maybe the optimizer fixes some of these usually?