Ricardo Vieira
Ricardo Vieira
### Description I would suggest moving those first six back to the `math` module scope. 
### Description When we use `pytensor.config.floatX=="float32"`, integer data is downcast to `"int16"` which has a pretty limited range of 32k. For count-based likelihoods this is way too narrow. I am...
### Description This is the default backend in nutpie and could possibly also be the one in PyMC. @aseyboldt suggests it may be much more performant than our numpy default...
Besides showing other samplers, we can also show the pymc samplers (prior, sample and posterior pred) using different backends with `compile_kwargs={"mode": "JAX"|"NUMBA"}`
The barebones PyTorch backend should be in in the next PyTensor release and subsequent PyMC. I think the best way to test it out and gather info on what Ops...
## TODO - [x] Implement QMC marginalization - [ ] Refactor away MarginalModel (separate PR) - [ ] Allow specifying desired Marginalization strategy - [ ] Wrap Sobol inside Op...
Brought up in #346 We can have a quick job that pip installs the package and tries to import it.
Closes #324 It's just `CategoricalGibbsMetropolis` but accepts `DiscreteMarkovChain` and knows how to get the dimensionality of P. Once we merge DMC into PyMC we can just extend the CategoricalGibbsMetropolis to...
Now that we have model transforms established in PyMC, I think we can do this via a model transform that eagerly places the `MarginalRV` is in the graph. Everything else...