I created a local test suite that tries to run bayes_kit.HMCDiag(..., stepsize=0.01, steps=50) and bayes_kit.MALA(epsilon=0.01) on suite of problems from posteriordb. I am just testing that outputs are not NaN and that no RuntimeErrors are encountered. I ran a total of 92 tests: the two algorithms above x 46 posteriors from posteriordb. 48/92 failed, and 44/92 passed.
This is not an issue with BayesKit per se. HMC and MALA can surely be made to work with smarter initialization and tuning. But this does call attention to the fact that it would be nice to support smarter initialization/warmup procedures here.
Also, I'm curious if it would be useful to contribute my posteriordb integration (including these tests) to this repository directly. If so, I can create a separate PR for that.
I created a local test suite that tries to run
bayes_kit.HMCDiag(..., stepsize=0.01, steps=50)andbayes_kit.MALA(epsilon=0.01)on suite of problems from posteriordb. I am just testing that outputs are not NaN and that noRuntimeErrors are encountered. I ran a total of 92 tests: the two algorithms above x 46 posteriors from posteriordb. 48/92 failed, and 44/92 passed.This is not an issue with BayesKit per se. HMC and MALA can surely be made to work with smarter initialization and tuning. But this does call attention to the fact that it would be nice to support smarter initialization/warmup procedures here.
Also, I'm curious if it would be useful to contribute my posteriordb integration (including these tests) to this repository directly. If so, I can create a separate PR for that.