Hi, my name is Rasul.
I’ve been reading the list of ideas for the GSoC 2020 and I would like to present my draft project proposal that I would like to implement this year: draft (dropbox). The proposal is not yet complete but the main points are listed, @junpenglao can you please review it and give some feedback on it.
Thank you.
This looks pretty good at a first glance. So your main proposal is subclassing tfp.mcmc so we can interact with it a bit more flexible?
Yeah, there are some ideas for design on subsampling samplers with parameterization. Also some issues raised that should be solved. I would love to here some additional ideas that could be added to the proposal.
@junpenglao can you please provide some feedback on the proposal draft, I’m still not sure if there is a need to support compound/Gibbs/progress bar and etc, so it is hard for me to provide some approximate timeline for every project idea. Also the PR on basic support for the basic support of these features is ready to review (design wise), there are still tests that should be added.
I will need a bit more time to review the code. As of the proposal, I suggest removing the progress bar part, but keep the compound/Gibbs step.
Ideally, we get a version that autogenerate state_fn for RWMH (so we can handle model that mixed with discrete dtype), and a metropolis-within-gibbs TransitionKernel (ie. compound step)
Alright, I guess I need to change couple of things in PR before then. There are some issues I see with the current implementation of bootstrap_results for compound step.
UPD: actually it will require more time to come up with the best solution to solve the current issues with trace handling for compound step, so I will add it to the proposal.