Prevent training without setting up caches.

[trivialfis]

• Add warning for internal functions.
• Check number of features.
  close [Blocking] python kernel failed when call Booster().predict #4056 .

[trivialfis]

Calling Booster.update without first initializing Booster with cache for dtrain results in num_feature == 0. I can raise an error in Python, but checking it in C++ seems more reliable. Added documentation effectively marked Booster.update and Booster.boost as internal methods.

@RAMitchell , @hcho3 Could you take a look see if there is better decision?

[codecov-io]

Codecov Report

Merging #4066 into master will decrease coverage by 0.01%.
The diff coverage is 0%.

@@            Coverage Diff             @@
##           master    #4066      +/-   ##
==========================================
- Coverage   60.56%   60.55%   -0.02%     
==========================================
  Files         130      130              
  Lines       11756    11758       +2     
==========================================
  Hits         7120     7120              
- Misses       4636     4638       +2

Impacted Files	Coverage Δ
python-package/xgboost/core.py	77.38% <ø> (ø)	⬆️
src/learner.cc	26.23% <0%> (-0.14%)	⬇️

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update 1fc37e4...5eb543a. Read the comment docs.

[RAMitchell]

Can we really say don't use something that is documented as a part of the public api? Maybe it is better to say 'power users only' or something.

[trivialfis]

Can we really say don't use something that is documented as a part of the public api? Maybe it is better to say 'power users only' or something.

Actually I don't like my solution. Let's see if I can make new datasets part of the caches safely.

[trivialfis]

@RAMitchell Turns out I can't make the incoming dataset for training to become part of the caches for the following reason:

In c_api, DMatrix handler is a std::shared_ptr, while Learner.UpdaterOneIter accepts raw pointer. Every time c_api::XGBoostUpdaterOneIter calls Learner::UpdaterOneIter, it first calls shared_ptr::get() to pass the raw pointer. So there's no way to let the caches from Learner obtain ownership of this DMatrix without changing Learner's interface, hence can not make it part of the caches (which is a vector of shared pointer).

And no, it's not "power user only". I'm a power user ( I think :) ), I don't know how to make it work other than making a copy of the abstracted APIs.

[trivialfis]

I will go ahead and merge this if no objections. @RAMitchell @hcho3

[hcho3]

Go ahead.

[trivialfis]

@thvasilo Thanks for the pointer.