Add keras mixin

[nateraw]

Adds mixin for keras

[Rocketknight1]

One question! What's the connection between this and the existing PushToHubMixin here? Are we deprecating that one in favour of this?

[nateraw]

One question! What's the connection between this and the existing PushToHubMixin here? Are we deprecating that one in favour of this?

@Rocketknight1 That's a hard-coded integration w/ transformers, AFAIK. In the future, it would probably be nice to use these Mixins directly instead and deprecate PushToHubMixin and other hub-related functions in transformers, but since these are changing a bit more rapidly, its probably for the best to leave the PushToHubMixin as-is for now.

[osanseviero]

Overall looks good, just a couple of questions

[LysandreJik]

Solid, I like what you've done with it! And the testing suite is very nice, too. LGTM!

[LysandreJik]

The kwargs are advertised as the kwargs to be passed to push_to_hub yet they're passed to _save_pretrained. I'd argue the kwargs to pass to each method ought to be different!

[LysandreJik]

My intuition tells me all of these should be optional kwargs, even if they're only supposed to be called internally

[osanseviero]

This looks great! Thanks a lot

[osanseviero]

Is this super class really needed? We only have one test class so maybe the API wrapper should be instantiated in setUpClass?

[nateraw]

You're probably right. I didn't want to mess with it too much, though. again, just following the previous file. I did find the superclass useful (if I remember right) when I played around with parametrizing the tests so we could run them over TF/PyTorch/etc.

[osanseviero]

Just a parenthesis discussion for a potential follow-up PR. Given that TF 2 suggested way of saving models is through SavedModels, should we also consider this use case? It's the recommended file format that is portable with TF Lite, TF Serving, etc.

Pros

• Much, much easier to use and save (easier to work with custom code, etc).
• Easy to connect with other tools of TF ecosystem (TensorFlow Serving for example)

Cons

• As discussed in the email, this was needed in transformers to only reload a part of a model from a checkpoint pre-trained with another architecture, which was only possible with the .h5 format.
• Probably other things I'm not aware of.

@nateraw, does the existing solution work with custom objects such as custom layers? Or will this only work if loaded with the same class? If not, I guess we can't use the created repos out of the box in the Inference API.

[nateraw]

@osanseviero I'm not sure if this supports custom layers/objects 😅 . I think we definitely want to use savedmodel going forward. This is a good first effort, but the next PR should be to use SavedModel. Didn't want to mess with this one too much just to keep things similar to whats available elsewhere in the HF ecosystem.