Refactor prediction

[alexpghayes]

This is an aggressive refactor to move all prediction code to safepredict. safepredict doesn't have methods for several of the existing models, so definitely don't merge until I've got that done. Plan is to push commits to this PR as an open project.

[alexpghayes]

I'm currently pulling most of the prediction tests into safepredict. That's where correctness checks should happen.

Once tests are passing, I'm planning to go back through and add simple tests not for correctness, but just that parsnip successfully passed the prediction off to safepredict and got a reasonable looking result in return. These sanity check tests still need to written, and probably should live in modeltests. I'm imagining something along the lines of

model <- fit(model_constructor(...), ...)
output <- predict(model, ...)
modeltests::check_predict_output(model, data, predict_options)

(also, this PR should be linked with #41)

[alexpghayes]

Current status: close to an MVP.

What I've done so far:

• Stripped out all the prediction code from parsnip
• Stripped out all the prediction tests from parsnip
• Moved it to safepredict

There's a fair amount of translation that needs to happen that is rather time consuming on the safepredict side. What I've got right now is minimal functionality -- most of the safepredict methods are thiiiiiin wrappers around predict(). They most definitely need to get fleshed out and tested on their own. Some (glm, glmnet, ranger) are much more fully featured, but still nothing is well tested.

Things I couldn't implement safe_predict methods for because of installation hell with sparklyr and keras:

• spark boosted trees
• spark linear model
• spark logistic reg
• spark random forest (class and reg)
• keras mlp regression
• keras mlp classification

These models will need safe_predict methods (they should have skeletons at the moment) and a sanity check that parsnip::predict() is getting something back.

I added some tiny tests that should check that at very least safe_predict is returning something with the right number of rows. So I think this PR won't cause any failures, but I could wrong. There may be some temporary feature loss as certain type arguments available in parsnip take a moment to become available in safepredict.

This PR has taught me an important lesson: namely, that it is sometimes better to make small incremental changes than to burn everything down and start from scratch.

[alexpghayes]

Oh, and I also need to add a kknn method to safe_predict() since Davis added kknn models.

Update: done.

[alexpghayes]

Ok this is so out of date I'm going to close and once safe_predict() is ready will re-open

[github-actions]

This pull request has been automatically locked. If you believe you have found a related problem, please file a new issue (with a reprex: https://reprex.tidyverse.org) and link to this issue.