This is weird…
The concatenate function works on the weight vector, more exactly it works to pack a Dirichlet distribution with a bunch of constant values (that I had to stick together with the “stack” method, by the way).
But the “concatenate” method output this error when used on the distribution of the mixture :
AsTensorError: ('Cannot convert <pymc3.distributions.continuous.Lognormal object at 0x7fbcc12c2438> to TensorType', <class 'pymc3.distributions.continuous.Lognormal'>)
I found it very hard to understand : all I want to do here is to pack one dimensional distributions (and constants) to build a heterogeneous mixture, but sometimes I need to use the “stack” function, sometimes the “concatenate”, and sometimes neither works…
I really don’t understand why what works for the weight vector distribution doesn’t work for the mixture components… Both are distributions!!
So, for the moment, my model can be executed, but is awfully slow and consume all the memory of my computer. I can only run it on a small dataset and reduced model dimension.
I guess , it is due to my last un-optimized loop. I really need to translate the python list of distributions mixture elements into a theano structure.
I tryied stack, stacklists, concatenate, but none works to stick the mixture elements together (which are log-normal, normal, and one halfnormal.)
Help of any kind will be appreciated.