Jeff Pollock

I'm not 100% sure as I don't use those layers, but I think you need to capture any batch dimensions in `t`: ``` python tfpl.DistributionLambda( make_distribution_fn=lambda t: tfd.Mixture( cat=tfd.Categorical(logits=[t[..., 0],...

+1 I would also like to know if `@info` etc should be supported inside `@floop` or what the recommended pattern is. Thanks!

I've also just noticed that `tfp.sts.Seasonal` doesn't have a parameter for the observation noise scale so really not sure what it is doing now - should it raise an error...

@davmre thanks that's really helpful! I'll have another try tomorrow and report back.

I've tried to re-organise the code a bit (mainly to help with my understanding) and run with TF1 but this also errors :-(. Here is the code: ``` python import...

Ah - thanks! Sorry I've been using TF 2 for a few weeks and totally forgot all about the global variable initializer. I have that running in about 8 seconds...

Hi @axch, thanks for the example - it is much nicer than mine. I’m travelling and not able to test anything at the moment, but I _think_ that this needs...

> This looks great! > > It'd be nice to see another test or two that the hessian approximation recovers (approximately) the covariance, and maybe a test that shows that...

> Hey @jeffpollock9 -- I used some of the ideas from your `_transform_reshape_split` bijector in the [new windowed sampler](https://github.com/tensorflow/probability/blob/master/tensorflow_probability/python/experimental/mcmc/windowed_sampling.py#L93) -- (re) using some of that code might be helpful here....

@ColCarroll if at all interesting, I've spent a little more time (not as much as I would like) on this and added another feature that I hope is useful -...