fix to make invert_hash work with bfgs (issue number 1891)

[agroh1]

Fixes issue where models trained with bfgs could not invert the hash.

I copied the code from ftrl.cc which does do this correctly.

There is also code in ftrl that does this for multipredict that I did not copy. I was not sure when that was used.

[jackgerrits]

It does seem like the other base reductions are doing this operation in their predict function too.

@JohnLangford @marco-rossi29 Are you aware of any differences in the way invert_hash is done between 8.5.0 and 8.6.0? Was it previously done in one of the GD::* calls that are being done here?

[JohnLangford]

invert_hash is essentially done via the audit pipeline. It's always functioned in this way.

[jackgerrits]

This block is present in the predict function of all base learners except bfgs though?

  if (audit)
      GD::print_audit_features(*(b.all), ec);

[JohnLangford]

This is an inconsistency.

…

-John

On Tue, May 28, 2019 at 10:00 AM Jack Gerrits ***@***.***> wrote: This block is present in the predict function of all base learners except bfgs though? if (audit) GD::print_audit_features(*(b.all), ec); — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub <#1892?email_source=notifications&email_token=AAATNZQQT3NZ7FZFR4OXSSLPXU3IBA5CNFSM4HPQQBZKYY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGODWMG5LI#issuecomment-496529069>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AAATNZRNGH5LUSS7NDQEFZDPXU3IBANCNFSM4HPQQBZA> .

[jackgerrits]

Sounds good, going to merge this in as it fixes that. Thanks @agroh1!