Skip to main content

Browse the archive

https://github.com/GPflow/GPflow
21 September 2021, 22:50:56 UTC
Revision 1164235071397edf4dc0e227992d6fae4bb3a5ed authored by Eric Hammy on 11 February 2020, 14:27:49 UTC, committed by GitHub on 11 February 2020, 14:27:49 UTC 
Inconisistency on whether q_sqrt is triangular fixed. (#1223)
 
* Inconisistency on whether q_sqrt is triangular fixed.

Problem:
* KL divergence calculations only take lower triangle parts of q_sqrt in
their calculation (implying a lower triangular parameterisation of
covariance)
* Conditional calculations perform calculations using the whole of
q_sqrt.
* The end result is if you forgot to add the triangular transform
to the q_sqrt, your forward pass may be the same (if you initialise with
a triangular matrix), but your gradients will immediately stop q_sqrt
being triangular.

Fix:
* Conditional calulcations now only take the lower triangle parts of
q_sqrt in their calculation
1 parent 7275533
Raw File
Tip revision: 1164235071397edf4dc0e227992d6fae4bb3a5ed authored by Eric Hammy on 11 February 2020, 14:27:49 UTC
Inconisistency on whether q_sqrt is triangular fixed. (#1223)
Tip revision: 1164235
VERSION 
2.0.0-rc1
back to top