-
Notifications
You must be signed in to change notification settings - Fork 229
Remove scale_identity_multiplier
to support tensorflow-probability 0.20
#828
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
Conversation
Updates the requirements on [tensorflow-probability](https://github.com/tensorflow/probability) to permit the latest version. - [Release notes](https://github.com/tensorflow/probability/releases) - [Commits](tensorflow/probability@0.8.0...v0.20.0) --- updated-dependencies: - dependency-name: tensorflow-probability dependency-type: direct:development ... Signed-off-by: dependabot[bot] <[email protected]>
Sanity check on
History with this PR:
|
Codecov Report
Additional details and impacted files@@ Coverage Diff @@
## master #828 +/- ##
==========================================
+ Coverage 81.16% 81.85% +0.68%
==========================================
Files 148 159 +11
Lines 9834 10321 +487
==========================================
+ Hits 7982 8448 +466
- Misses 1852 1873 +21
|
I think it's fine to have this change be breaking - it would be more unexpected behaviour if we kept it to Would be good to check all call-sites in the library and on the docs that utilize the |
Thanks @jklaise. Re our library callsites for alibi-detect/alibi_detect/od/vae.py Line 81 in 8b2dcbb
But we're already setting |
In
tensorflow-probability
v0.20 thescale_identity_multiplier
kwarg given toMultivariateNormalDiag
is removed, meaning the covariance matrix must be specified viascale_diag
(see tensorflow/probability@ca84850)This PR modifies our
elbo
loss function, so thatscale_diag
is constructed from thesim
kwarg, rather than passing it directly toscale_identity_multiplier
. Currently, only one ofcov_full
,cov_diag
orsim
should be given toelbo
.Breaking change: This currently involves a breaking change since the default value of
sim
inelbo
has been changed fromsim=0.05
tosim=None
. A possible alternative would be to havesim=0.05
as the default, and then setsim=None
under-the-hood when eithercov_diag
orcov_full
are given (i.e.not None
).