in summary, from gitter history, pymc3 is ‘stuck’ with theano. as far as functionality, theano is missing proper multiple compute device support (gpus & cpus located anywhere in a network).
can anyone talk about a use case where more computing than can be found in a single machine would be useful?
Yeah we rely on a lot of Theano magic at the moment, switching to tensorflow would be a major breaking change.
I think the advantage of the current implementation in Theano is that for smaller model it is much faster than Tensorflow (from the @aseyboldt’s experiment). You can try out different models much easier. But it is true that if you want to deploy a model into production and/or use GPU, parallel processing things become much more difficult.
i expect that advantage to diminish with improvements to tensorflow. one feature that’s being developed is compiling tf graphs. that should make things faster.
pymc4 anyone?