Techniques like SMC, Langevin sampling and MALA don't seem to be covered by numpyro out of the box. I am aware that using blackjax or TFP with an adapter is an option in some cases, but still feel that a numpyro-native implementation could lower the barrier to using a wider range of samplers and make modelling code more concise and self-contained. Would you be open to accepting contributions that implement these missing samplers? If positive in principle, very happy to discuss case by case here or in separate issues.
Techniques like SMC, Langevin sampling and MALA don't seem to be covered by numpyro out of the box. I am aware that using blackjax or TFP with an adapter is an option in some cases, but still feel that a numpyro-native implementation could lower the barrier to using a wider range of samplers and make modelling code more concise and self-contained. Would you be open to accepting contributions that implement these missing samplers? If positive in principle, very happy to discuss case by case here or in separate issues.