Pre-warmup automatic reparameterization?

[Mike Lawrence]

Has the approach of Gorinova et al (2019) made it into TFP in any form? I'd love to cross centered/non-centered tuning off my list of model-building decisions!

[Mike Lawrence]

Oh, I see that at least the numpyro folks implemented it...

[Christopher Suter]

Hi Mike, we don't have anything pre-packaged (that I'm aware of) for the parametric semi-decentering, but here's a colab demo of non-centering Neal's funnel using `tfp.experimental.bijectors.make_distribution_bijector` (generously bestowed upon us by Dave Moore, co-author w/ Maria Gorinova on the paper you cited):

https://colab.research.google.com/drive/1gqrMbcYljYkeLS1d2EdvjODt9Anx2tIs?usp=sharing

There are more examples in the docstring for make_distribution_bijector. It can do some pretty fancy stuff.

On Thu, Apr 7, 2022 at 4:20 PM Mike Lawrence <mike....@gmail.com> wrote:

Oh, I see that at least the numpyro folks implemented it...

--
You received this message because you are subscribed to the Google Groups "TensorFlow Probability" group.
To unsubscribe from this group and stop receiving emails from it, send an email to tfprobabilit...@tensorflow.org.
To view this discussion on the web visit https://groups.google.com/a/tensorflow.org/d/msgid/tfprobability/CAB%2BQPJCDTCTLgtBmdf7GZ8oAL_ijQCUimCN-Gccm%2BmKmxL4iww%40mail.gmail.com.

[Pavel Sountsov]

The way you parameterize (I think) it is you first train a structured surrogate posterior via VI (see an example in the build_asvi_surrogate_posterior docstring) and then use `tfp.experimental.bijectors.make_distribution_bijector` on the learned surrogate.

To view this discussion on the web visit https://groups.google.com/a/tensorflow.org/d/msgid/tfprobability/CACaB3_2sREy0hyW61dZOj76%2Bgvgk0K_73wT0NpOdvx-R7Vu3gw%40mail.gmail.com.