BUG Initialize adapt_full with a diagonal covariance matrix rather than ones.

[twiecki]

No description provided.

[eigenfoo]

Oof, not sure how that slipped by me the first time! This looks good: will merge once tests pass.

[fonnesbeck]

Oops. Good catch.

[codecov]

Codecov Report

Merging #3891 into master will not change coverage.
The diff coverage is 0.00%.

@@           Coverage Diff           @@
##           master    #3891   +/-   ##
=======================================
  Coverage   90.71%   90.71%           
=======================================
  Files         135      135           
  Lines       21317    21317           
=======================================
  Hits        19338    19338           
  Misses       1979     1979           

Impacted Files	Coverage Δ
pymc3/sampling.py	84.52% <0.00%> (ø)

[fonnesbeck]

Not sure if its a coincidence, but I've been getting sampling failures like this after updating:

ValueError: Mass matrix contains zeros on the diagonal. 
The derivative of RV `sigma_log__`.ravel()[0] is zero.
The derivative of RV `sigma_log__`.ravel()[1] is zero.
The derivative of RV `sigma_log__`.ravel()[2] is zero.
The derivative of RV `sigma_log__`.ravel()[3] is zero.
The derivative of RV `sigma_log__`.ravel()[4] is zero.
The derivative of RV `sigma_log__`.ravel()[5] is zero.
The derivative of RV `sigma_log__`.ravel()[6] is zero.
The derivative of RV `sigma_log__`.ravel()[7] is zero.

[twiecki]

Using this init?

[fonnesbeck]

Yeah, but now it seems to be happening for other init methods, so it must have been a coincidence. Please ignore.

[ColCarroll]

Still might want to init this like the diagonal + jitter (i.e., use the mean plus some random noise -- shoudl be a copy/paste)

[ColCarroll]

I can get to it later if no one else does!

[twiecki]

Great!

…

On Mon, Apr 27, 2020, 22:49 Colin ***@***.***> wrote: I can get to it later if no one else does! — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub <#3891 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AAFETGBFUKREAZYQUOFTORLROXVVZANCNFSM4MSBMSFQ> .