Improve with providing Nuclear Norm Constraint

[pkan2]

Hello!

I have updated the conditional_gradient optimizer with providing the option of nuclear norm constraint option, as described in issue #1105,

with using following formula:

variable -= (1-learning_rate) * (variable
    + lambda_ * top_singular_vector(gradient))

Thank you!

[pkan2]

Hello @facaiy @WindQAQ @Squadrick!

Can you provide some feedback for the update of CG optimizer?

Thank you!

[Squadrick]

Most of the logic for fro and nuclear norm in _resource_apply_dense and _resource_apply_sparse is the same except for var_slice = tf.gather(...). Refactoring the logic into functions would make the code more succinct and easier to read.

Yet to properly review the test code.

[gabrieldemarmiesse]

@googlebot I consent

[googlebot]

CLAs look good, thanks!

ℹ️ Googlers: Go here for more info.

[gabrieldemarmiesse]

I've merged master into your branch to update it and fixed any formatting/conflicts it might have. If you need to do some more modifications, please do git pull beforehand.

[pkan2]

@gabrieldemarmiesse Got it! Thank you a lot!

[gabrieldemarmiesse]

@Squadrick when you have the time, could you look into this pull request? Thank you.

[Squadrick]

@pkan2 There's a merge conflict, can you fix that? Once it's done, this should get merged.

[pkan2]

Hello @Squadrick . Thank you for your feedback! I have fixed the merging conflict.

Thank you!

[Squadrick]

LGTM.

[gabrieldemarmiesse]

I'll merge this as it was approved by @Squadrick