InexactError: trunc(UInt8, 256) when user sets nbins>255

[svilupp]

Current behaviour: If a user sets nbins>255 they will get InexactError: trunc(UInt8, 256)

Expected behaviour: Either the restriction should be explicit and caught on initialization or the code should be updated.

Possible cause:

• Once features are split, all data points are bucketed into corresponding slots by function binarize
• This line makes an assumption that the buckets are of type UInt8, so implicitly assuming there are never more than 255 bins

Suggested solution:

• 255 bins should be sufficient and no user has run into this limitation so far, so it can be kept
• we should add a note in the documentation
• and check this value on initialization (with a helpful error message)

I'm happy to take a stab at the PR if helpful.

[svilupp]

MWE (taken from the example in docs):

using EvoTrees
using Random

x_train = randn(100, 10)
y_train = randn(100)

config = EvoTreeRegressor(
    loss=:linear, 
    nrounds=100, 
    max_depth=6, 
    nbins=256,
    eta=0.1,
    lambda=0.1, 
    gamma=0.1, 
    min_weight=1.0,
    rowsample=0.5, 
    colsample=0.8)

m = fit_evotree(config; x_train, y_train)
preds = m(x_train)

# throws: nested task error: InexactError: trunc(UInt8, 256)

Tested on the latest version in main. versioninfo()

Julia Version 1.8.5 Commit 17cfb8e65ea (2023-01-08 06:45 UTC) Platform Info: OS: macOS (arm64-apple-darwin21.5.0) CPU: 8 × Apple M1 Pro WORD_SIZE: 64 LIBM: libopenlibm LLVM: libLLVM-13.0.1 (ORCJIT, apple-m1) Threads: 6 on 6 virtual cores

[jeremiedb]

Adding domain validation of the arguments passed to the hyper-params constructors is something I've been shoveling forward a bit too long! Thanks for reporting, feel free to go for a PR, otherwise, I'll likely get it done with 1-2 days.

[svilupp]

I have too many other things going on this weekend, but I should be able to get to it on Monday.

When I find time, I'll check here if you've started or not.

[svilupp]

I've hacked up a draft. Let me know what you think!