[BUG] cuML conda package - cuml.svm.SVC depends on sklearn

[ksangeek]

Describe the bug
cuml.svm.SVC depends on sklearn, but the conda package doesn't have scikit-learn as a run dependency.

Steps/Code to reproduce bug

1. I installed cuml 0.11.0 in a new python 3.6 conda environment using -

conda install -c rapidsai -c nvidia -c conda-forge -c defaults cuml=0.11 python=3.6 cudatoolkit=10.1

2. Using SVC fails with -

python
Python 3.6.10 |Anaconda, Inc.| (default, Jan  7 2020, 21:14:29)
[GCC 7.3.0] on linux
Type "help", "copyright", "credits" or "license" for more information.
>>> from cuml.svm import SVC
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "/home/sangeek/anaconda3/envs/cuml-11-py36/lib/python3.6/site-packages/cuml/svm/__init__.py", line 16, in <module>
    from cuml.svm.svm import SVC
  File "cuml/svm/svm.pyx", line 36, in init cuml.svm.svm
ModuleNotFoundError: No module named 'sklearn'

Expected behavior
>>> from cuml.svm import SVC should work without errors
It is very clear that installing scikit-learn in the conda environment would fix the issue as SVC requires it -

cuml/python/cuml/svm/svc.pyx

Line 36 in 6fbba53

from sklearn.exceptions import NotFittedError

And the only change to fix this is to use scikit-learn as a run dependency in the conda packaging here -

cuml/conda/recipes/cuml/meta.yaml

Line 38 in 6fbba53

run:

Environment details (please complete the following information):

• Environment location: [Bare-metal]
• Linux Distro/Architecture: [RHEL]
• GPU Model/Driver: [V100 and driver 396.44]
• CUDA: [10.1]
• Method of cuDF & cuML install: [conda]
  • If method of install is [conda], run conda list and include results here
  • If method of install is [Docker], provide docker pull & docker run commands used
  • If method of install is [from source], provide versions of cmake & gcc/g++ and commit hash of build

Additional context
Add any other context about the problem here.

[dantegd]

@ksangeek thanks for catching this! It was an oversight, we have removed the runtime requirement (which is overkill for one single exception type, particularly when we're working on our own exceptions: #817 )