Dear PyMC3 community,
We are trying to infer the effectivity of social distancing measures on the spread of Covid-19 and are looking for support. If you are interested look at our repository. We can provide funding on PostDoc, Phd and master student level: https://github.com/Priesemann-Group/covid19_inference/
We made some nice plots for Germany: https://github.com/Priesemann-Group/covid19_inference_forecast
By the way, our analysis has been published now in Science! https://science.sciencemag.org/content/early/2020/05/14/science.abb9789
Thanks to all contributors of PyMC3. Especially all the explanatory notebooks/tutorials were very helpful.
Great work! Congratulations!
I just discovered the notebook Estimating_COVID_19_in_11_European_countries, in which you also seem to be involved, and the corresponding paper it is based on. This is fascinating, this is exactly one of the directions in which we wanted to continue our research, and I think you already solved at least one of the problems that I am having right now with the heaviness of tails of some lognormal distributions by using a tfb.SinhArcsinh bijection. I am thinking right now whether to switch to tensorflow probability. My hierarchical models in PyMC3 are also running 1-2 hours and a running time in the range of minutes would help a lot…
Oh fantastic! Yes fine tuning bijectors is actually the thing that makes inference fast and numerically more stable, the default Softplus or Exp perform much worse.
@jdehning Also, the TFP team is preparing to release some additional code on this model to demonstrate alternative inference method and computation hardware (TPU). For example, you can run a SMC to get a good idea of the posterior distribution in a few minutes (without worrying too much of bijectors although a good bijection also helps a lot here). I will update this thread once that’s available.