Input/Output Feature Schema Refactor

[connor-mccorm]

Adds Marshmallow schema support for all input and output features and their corresponding preprocessing, encoders, and decoders.

[github-actions]

Unit Test Results

       6 files  ±  0         6 suites  ±0   2h 24m 3s ⏱️ + 2m 7s
2 919 tests +  4  2 873 ✔️ +  4    46 💤 ±0  0 ❌ ±0 
8 757 runs  +12  8 615 ✔️ +12  142 💤 ±0  0 ❌ ±0 

Results for commit cd43485. ± Comparison against base commit dfdc98c.

♻️ This comment has been updated with latest results.

[connor-mccorm]

Most tests are passing, there is one case where I need to add a regex pattern optional argument for the string input type to handle custom fill value restrictions, but once I get that done, we should be good to go.

[martindavis]

Hey @connor-mccorm , overall excellent job making the Schema docs very clean and clear in Ludwig! I've got some questions, but I don't see anything blocking this from my perspective.

[tgaddair]

Is the plan to wire this into the schema validation logic in a follow-up?

[connor-mccorm]

So I believe the schema validation logic will be using all this code now when this lands through getting each input feature marshmallow class via a get_schema_cls() method in each input feature type. The main caveat here is that the defaults for input feature preprocessing are still being set with the preprocessing_defaults() method of each input feature type class since it looked like there we're a handful of dependencies that would need refactoring. So I plan on fixing that in a followup PR asap. Does that sound ok to you @tgaddair?

[tgaddair]

Sounds good @connor-mccorm !