This PR turned out to be trickier and larger than expected. Anyway, here it goes with the following changes:
We now use auxiliary workflow for our typical benchmark rules. Note that it now scales tasks automatically, so that UNSCALED_TASKS should be declared rather than TASKS. In particular, look how much cool_tasks.smk has shrunk.
It adds a product distribution (for independent random variables) to the fine subpackage. It also adds to the public API utilities for creating TensorFlow Probability's multivariate normal and Student distributions. (Which were previously hidden).
It adds the workflow for studying outliers/failing communication channels. It's ugly, but it seems that it does the job.
If you think it's ready to be merged, please merge – I will have little coding time until Wednesday evening...
This PR turned out to be trickier and larger than expected. Anyway, here it goes with the following changes:
UNSCALED_TASKSshould be declared rather thanTASKS. In particular, look how muchcool_tasks.smkhas shrunk.finesubpackage. It also adds to the public API utilities for creating TensorFlow Probability's multivariate normal and Student distributions. (Which were previously hidden).If you think it's ready to be merged, please merge – I will have little coding time until Wednesday evening...