tfp.mcmc.NoUTurnSampler is not working with ReplicaExchangeMC

Damian Ndiwago

unread,

Mar 23, 2022, 4:50:20 PM3/23/22

to TensorFlow Probability

Hi all,

I get error REMCFieldNotFoundError: NUTSKernelResults( ) or a kernel result nested within it is currently not supported by the ReplicaExchangeMC kernel. Is there any way I can get around this for NUTS? Here is a link to the example I run

https://colab.research.google.com/drive/1QQYIbH0-lWcxJGd2M5mwpAO-XDuwoWA3?usp=sharing. I appreciate any help.

Thank you,

Damian

Junpeng Lao

unread,

Mar 26, 2022, 4:51:13 AM3/26/22

to TensorFlow Probability, ndiwag...@gmail.com

Hi Damian,

I am not sure it is possible to use NUTS with ReplicaExchangeMC, as NUTS proposal contains multiple proposals (any state on the Hamiltonian trajectory during the binary tree building are valid proposal as they are on the same level set), unlike HMC or MH that contains a single proposal, which the information related to the proposal is used by ReplicaExchangeMC.

Damian Ndiwago

unread,

Mar 30, 2022, 3:02:24 AM3/30/22

to Junpeng Lao, TensorFlow Probability

Hi Junpeng,

Thank you so much, I get it now.

Regards,

Damian

Reply all

Reply to author

Forward