Crash in `RandomForestClassifier.fit` when passing in continuous target

[jcrist]

In #6863 a test in the sklearn test suite was identified that causes RandomForestClassifier.fit to crash:

cuml/python/cuml/cuml/accel/tests/scikit-learn/xfail-list.yaml

Line 1081 in 05426a0

- "sklearn.tests.test_common::test_estimators[RandomForestClassifier()-check_classifiers_regression_target]"

This test tries passing a continuous target (e.g. a regression dataset) to a classifier, and expects the classifier to error appropriately notifying the user of their mistake.

Here's a reproducer with just cuml core:

from sklearn.datasets import make_regression
from cuml import RandomForestClassifier

X, y = make_regression(
    n_samples=200,
    n_features=10,
    n_informative=1,
    bias=5.0,
    noise=20,
    random_state=42,
)

model = RandomForestClassifier()
model.fit(X, y)

Output

$ python bug.py
terminate called after throwing an instance of 'raft::exception'
  what():  exception occurred! file=/raid/jcristharif/cuml/cpp/src/decisiontree/batched-levelalgo/builder.cuh line=513: Not enough shared memory. Consider reducing max_n_bins.
Obtained 7 stack frames
#1 in /raid/jcristharif/miniforge3/envs/cuml-dev/lib/libcuml++.so: ML::DT::Builder<ML::DT::GiniObjectiveFunction<double, int, int> >::computeSplit(int, unsigned long, unsigned long) +0x239 [0x7f5b41d617d9]
#2 in /raid/jcristharif/miniforge3/envs/cuml-dev/lib/libcuml++.so: ML::DT::Builder<ML::DT::GiniObjectiveFunction<double, int, int> >::doSplit(std::vector<ML::DT::NodeWorkItem, std::allocator<ML::DT::NodeWorkItem> > const&) +0x281 [0x7f5b41d77691]
#3 in /raid/jcristharif/miniforge3/envs/cuml-dev/lib/libcuml++.so: ML::DT::Builder<ML::DT::GiniObjectiveFunction<double, int, int> >::train() +0x4cc [0x7f5b41d8b66c]
#4 in /raid/jcristharif/miniforge3/envs/cuml-dev/lib/libcuml++.so(+0x88e929) [0x7f5b41d8d929]
#5 in /raid/jcristharif/miniforge3/envs/cuml-dev/lib/libgomp.so.1(+0x19ec4) [0x7f5b8c513ec4]
#6 in /lib/x86_64-linux-gnu/libpthread.so.0(+0x8609) [0x7f5eff534609]
#7 in /lib/x86_64-linux-gnu/libc.so.6: clone +0x43 [0x7f5eff2ff353]

Aborted (core dumped)

[Ameya674]

Hi, I’d like to pick this up—can I be assigned?

[csadorf]

Hi, I’d like to pick this up—can I be assigned?

Contributions are very welcome, thank you!

[csadorf]

@Ameya674 were you able to make any progress on this?

[Ameya674]

Hey, still working on it, will hopefully finish it by the week...

[csadorf]

Ok! We would need a PR ready within the next few days if the fix should make it reliably into the next release. CC: @divyegala

[betatim]

A fix for this is to add a call to (something like) sklearn.utils.multiclass.check_classification_targets on or after

cuml/python/cuml/cuml/ensemble/randomforest_common.pyx

Line 395 in ff1c1af

if self.RF_type == CLASSIFICATION:

I'm not sure if we can use the function directly or if we'd have to implement our own. Most of the work happens in type_of_target https://github.com/scikit-learn/scikit-learn/blob/4a4e5f52726aa5c2b21a2cda326a8be921271611/sklearn/utils/multiclass.py#L228