Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Parameters returned by the warmup are off #116

Closed
rlouf opened this issue Sep 17, 2021 · 4 comments · Fixed by #218
Closed

Parameters returned by the warmup are off #116

rlouf opened this issue Sep 17, 2021 · 4 comments · Fixed by #218
Labels
adaptation Issue related to the adaptation bug Something isn't working important
Milestone

Comments

@rlouf
Copy link
Member

rlouf commented Sep 17, 2021

See also #44 for a failure on a simple gaussian target. Issue was spotted while addressing #112.

Bug Description

In the Introduction.ipynb example, the kernel built using parameters obtained with the window adaptation results in only divergences due to a value for the step size that is too large.

Versions

Python implementation: CPython
Python version       : 3.9.6
IPython version      : 7.22.0

jax     : 0.2.20
jaxlib  : 0.1.71
blackjax: 0.2.x

Compiler    : GCC 11.1.0
OS          : Linux
Release     : 5.13.13-arch1-1
Machine     : x86_64
Processor   : 
CPU cores   : 24
Architecture: 64bit
@rlouf rlouf added bug Something isn't working important labels Sep 17, 2021
@rlouf rlouf added this to the 0.3.0 milestone Sep 17, 2021
@rlouf
Copy link
Member Author

rlouf commented Sep 20, 2021

Note: this PR in PyMC3 improves the warmup by waiting for the chain to be in the typical set before adapting the mass matrix.

@rlouf
Copy link
Member Author

rlouf commented Sep 29, 2021

Note: simplify the implementation of the dual averaging algorithm as I did in aehmc.

@rlouf
Copy link
Member Author

rlouf commented Nov 17, 2021

Looks like the acceptance probability currently returned is not the expected value, which may explain the problem with the warmup.

@junpenglao junpenglao linked a pull request Feb 8, 2022 that will close this issue
@rlouf rlouf linked a pull request Jun 10, 2022 that will close this issue
@rlouf rlouf modified the milestones: 0.3.0, 1.0 Jun 29, 2022
@rlouf rlouf added the adaptation Issue related to the adaptation label Jul 3, 2022
@rlouf
Copy link
Member Author

rlouf commented Jul 3, 2022

The Introduction.ipynb notebook was the last place where there was an issue with adaptation (to my knowledge). I updated the notebook in a29272d and found that the resulting kernel samples correctly 👍

@rlouf rlouf closed this as completed Jul 3, 2022
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
adaptation Issue related to the adaptation bug Something isn't working important
Projects
None yet
1 participant