This poll will help us gather information about the different settings in which people are using Pyro PPL. All questions are multiple choice.
If you wish to provide more context or more nuanced answers, please feel free to add a free text response below.
In which context(s) do you use Pyro/NumPyro?
- Applied Research
- Algorithms Development Research
- Personal Projects
What PPLs do you use?
Which Pyro/NumPyro inference algorithm(s) do you use?
If you use SVI, what kind of guides do you use?
- Auto Guides
- Custom Guides
- I do not use SVI (see results)
Which of the following model classes do you typically implement using Pyro/NumPyro?
- Traditional Bayesian statistics models (i.e. no neural networks)
- Machine learning models (variational autoencoders etc.)
Which device accelerators do you use for inference with Pyro/NumPyro?
I personally find myself using Pyro way more than Numpyro just because it is based on PyTorch, which allows me to create complex models based on PyTorch’s library ecosystem (mostly combining Pyro with gpytorch and botorch to train complex gaussian processes using Pyro’s SVI routines).
I would love to see some support for Pyro contrib’s
HiddenLayer class in order to have again proper bayesian neural network support in Pyro.
Thank you for your great work in (Num)Pyro!