[Plugin] Keras types integration

[cosmicBboy]

Flyte is an ML- and data- aware orchestrator, it would make sense to support the most popular ML frameworks as first-class citizen types in the Flyte type system.

One of the popular deep learning frameworks is keras, which would make sense to support without installing plugins and simply be activated when a user has it installed, similar to how we're handling pytorch, tensorflow (#2569, #2570, #2571), and sklearn (#2752).

We'll need to follow the conventions set in the flytekit.extras.pytorch module, by doing the following:

1. the __init__.py file should handle the case where the user doesn't have keras installed.
2. implement a TypeTransformer for the following classes:
   • Model: https://keras.io/api/models/model/
   • Sequential: https://keras.io/api/models/sequential/
   • (TODO: See if there's a single parent class for all keras models 🤷‍♂️)
3. add tests here: https://github.com/flyteorg/flytekit/tree/master/tests/flytekit/unit/extras

Resources

• flyte docs: https://docs.flyte.org/projects/cookbook/en/latest/auto/core/extend_flyte/custom_types.html
• saving/loading keras models: https://keras.io/api/models/model_saving_apis/

[SmritiSatyanV]

Could I work on this? @cosmicBboy

[ryankarlos]

@samhita-alla ive added a PR for this

[ryankarlos]

#take

[dennisobrien]

There is another issue that is probably a duplicate of this. #2570
And that issue also has a PR in progress. flyteorg/flytekit#1241

[samhita-alla]

@dennisobrien, you're right. @cosmicBboy, which PR shall we close?

[ryankarlos]

@samhita-alla Ive also added support for Sequential class in this PR - i could port that over to the tensorflow folder and associated scripts once the duplicate PR in flyteorg/flytekit#1241 is merged in, if that helps.
Not sure if support for type transformer for https://keras.io/api/layers/ would be useful as well but happy to add that in too.