Add a warning when fitting models loaded from SavedModels

[bhack]

Add a warning when fitting models loaded from SaveModel having optimizers with slots.
Ref: #42749 (comment)

[bhack]

/cc @reedwm @k-w-w

[bhack]

This is just an exploring proposal. If you have a better entry point let me know.

[gbaned]

@bhack @reedwm @k-w-w This PR is in draft, any update on this? Please. Thanks!

[bhack]

@gbaned I have any private channel so the thread it is what you see here.

[reedwm]

Can you give context for this PR? I think adding a warning in fit() for loaded saved models with slot variables is sensible, but I don't see any warnings added in this PR.

[bhack]

@reedwm If you see is a draft as I have asked for comment on 1st September about the entry point I selected in the code to "introduce" the condition but as of 2nd December I got any feedback.
I have any other chat channel to discuss with devs as Gitter SIG Keras-io/Lobby in the current status is useless for discussion about PRs.

[reedwm]

Unfortunately I am not very familiar with the SavedModel code so I'm not sure if this is a good place to do it. Sorry for not responding until now.

@k-w-w, do you think it would be a good idea to add a warning if a SavedModel with slot variables is restored and Model.fit() is used? Currently slot variables are not restored in such a case, which could affect training and is why we should potentially add a warning. If this is a good idea, would the code modified in this PR be a good place to check for the use of slot variables?

[bhack]

@reedwm Thank you. Just to note that I've already asked to @k-w-w on 23th Oct at tensorflow/community#281 (comment)

[k-w-w]

The change looks good, although I think directly sending the warning when loading would make sense (see the warning message for the HDF5 format:

tensorflow/tensorflow/python/keras/saving/hdf5_format.py

Line 211 in a6b7e4b

'your model is starting with a freshly initialized optimizer.')

)

[bhack]

The change looks good, although I think directly sending the warning when loading would make sense (see the warning message for the HDF5 format:

tensorflow/tensorflow/python/keras/saving/hdf5_format.py

Line 211 in a6b7e4b

'your model is starting with a freshly initialized optimizer.')

)

Do you like the same HDF5 warning message or something else?

[bhack]

Ok Windows is failing for unrelated MLIR issues and Macos CPU is failing probably for other unrelated stuff but has any visible details.
The other checks are passing.

[k-w-w]

Sorry for the delay, thank you!

[bhack]

@k-w-w The sanity is temp broken by #48294 (comment)

[k-w-w]

@k-w-w The sanity is temp broken by #48294 (comment)

Oh I see, good to know

[bhack]

Can you re-trigger the build? I think that the CI Sanity it is working again.

[bhack]

Do you need anything else here?

[k-w-w]

Hmm I'm not sure why this isn't being automatically pulled in. Let me try re-approving.

[bhack]

@k-w-w If you want you can add a reference somewhere to #44670