Adrian Seyboldt
Adrian Seyboldt
I mostly like to use it to identify correlations and divergences, and for those I think it is a lot more convenient to have them in a 1d array: https://pymc-devs.github.io/nutpie/sample-stats.html#covariance-of-gradients-and-draws
Yes, this is definitely something we want to support. I did not yet expose it as a public function because I thought I might want to do some minor changes...
I agree that this would be nice to have (and it wouldn't be that hard to implement, it only needs some changes in the `compile_pymc_model` function. If someone wants to...
What's happening is the following: nutpie calls into bridgestan to convert the stan model code to c++ and compile this using code from the stan repository. The c++ code depends...
Ah, that explains why this doesn't happen for cmdstan. A conda bridgestan package would indeed be nice :-) For the snipped, I don't think you would have to compile anything...
Coming up soon :-)
Current work is here: https://github.com/pymc-devs/nutpie/pull/144
Since I don't think we can do much about the numba compilation times in nutpie, I'll close this here.
Sorry for the slow reply. You are right, the warmup draws are only discarded at the very end, so not storing them does currently not reduce the maximum memory requirement....
Version 0.16 includes a zarr backend, so that you don't have to keep the whole trace in memory. It also lowers the memory requirement without the zarr backend a little...