Normalizing Flow MCMC

[Qazalbash]

Description

It would be great to have Normalizing Flow MCMC method.

Paper reference

• Adaptive Monte Carlo augmented with normalizing flows
• Efficient Bayesian Sampling Using Normalizing Flows to Assist Markov Chain Monte Carlo Methods

[fehiepsi]

Hi @Qazalbash, please check out https://github.com/kazewong/flowMC . We do not have plan to support those algirthms.

[renecotyfanboy]

I think you can achieve pretty similar results using the flow reparametrizations that are already implemented : https://num.pyro.ai/en/latest/examples/neutra.html

[h2o64]

I think you can achieve pretty similar results using the flow reparametrizations that are already implemented : https://num.pyro.ai/en/latest/examples/neutra.html

Hi @renecotyfanboy I released a paper last year answering this very question. FlowMC is significantly superior to NeuTra on multimodal distributions. I believe its addition would be nice.

[renecotyfanboy]

That's super interesting, I had no idea there would be significant differences between flow reparametrization and proposal. This could be a great addition in the contributed codebase if there is no plan to maintain it in the core package

[fehiepsi]

I think it is better to use inference utilities on numpyro models to get the joint density and use flowMC on the top of that. It would be great to have a tutorial for it.

[fehiepsi]

Please checkout https://num.pyro.ai/en/latest/tutorials/other_samplers.html for integration with other apis.