[BUG] Non-linear SVC and SVR are incompatible with dask-ml GridsearchCV

[beckernick]

Referencing the coef_ attribute on SVC and SVR throws a generic RuntimeError rather than an AttributeError for non-linear kernels. This breaks dask-ML's GridsearchCV (and potentially other components), as it relies on explicit error catching in conditional logic: https://github.com/dask/dask-ml/blob/f1933bbd344452d8cec900ca902da4e417e1683e/dask_ml/model_selection/_normalize.py#L37-L40

Note that this still breaks even if the parameter dict passed to dask-ml only specifies the linear kernel. As the default kernel is rbf, you'd need to also explicitly specify this in the estimator passed to GridsearchCV.

        for attr in attrs:
            if attr in exclude:
                continue
            try:
                val = getattr(est, attr)
            except (sklearn.exceptions.NotFittedError, AttributeError):
                continue

If possible, we should try to throw the more specific AttributeError in this scenario.

from sklearn.svm import SVC, SVR
import cuml
​
getattr(SVR(), "coef_")
---------------------------------------------------------------------------
AttributeError                            Traceback (most recent call last)
/tmp/ipykernel_55047/1241473841.py in <module>
      2 import cuml
      3 
----> 4 getattr(SVR(), "coef_")

/raid/nicholasb/miniconda3/envs/rapids-21.10/lib/python3.8/site-packages/sklearn/svm/_base.py in coef_(self)
    499     def coef_(self):
    500         if self.kernel != 'linear':
--> 501             raise AttributeError('coef_ is only available when using a '
    502                                  'linear kernel')
    503 

AttributeError: coef_ is only available when using a linear kernel

from sklearn.svm import SVC, SVR
import cuml
​
getattr(cuml.svm.SVR(), "coef_")
---------------------------------------------------------------------------
RuntimeError                              Traceback (most recent call last)
/tmp/ipykernel_55047/3521656981.py in <module>
      2 import cuml
      3 
----> 4 getattr(cuml.svm.SVR(), "coef_")

/raid/nicholasb/miniconda3/envs/rapids-21.10/lib/python3.8/site-packages/cuml/internals/api_decorators.py in inner(*args, **kwargs)
    593 
    594                 # Call the function
--> 595                 ret_val = func(*args, **kwargs)
    596 
    597             return cm.process_return(ret_val)

cuml/svm/svm_base.pyx in cuml.svm.svm_base.SVMBase.coef_()

RuntimeError: coef_ is only available for linear kernels

Dask-ML Failing Example (the error is propagated up rather than caught):

import dask_ml.model_selection as dcv
from sklearn.datasets import load_iris
import cuml
​
iris = load_iris()
​
parameters = {'kernel': ['rbf'], 'C': [1, 10]}
svc = cuml.svm.SVC()
clf = dcv.GridSearchCV(svc, parameters)
clf.fit(iris.data, iris.target)
---------------------------------------------------------------------------
RuntimeError                              Traceback (most recent call last)
/tmp/ipykernel_55610/1025904613.py in <module>
      8 svc = cuml.svm.SVC()
      9 clf = dcv.GridSearchCV(svc, parameters)
---> 10 clf.fit(iris.data, iris.target)

/raid/nicholasb/miniconda3/envs/rapids-21.10/lib/python3.8/site-packages/dask_ml/model_selection/_search.py in fit(self, X, y, groups, **fit_params)
   1234 
   1235         candidate_params = list(self._get_param_iterator())
-> 1236         dsk, keys, n_splits, _ = build_cv_graph(
   1237             estimator,
   1238             self.cv,

/raid/nicholasb/miniconda3/envs/rapids-21.10/lib/python3.8/site-packages/dask_ml/model_selection/_search.py in build_cv_graph(estimator, cv, scorer, candidate_params, X, y, groups, fit_params, iid, error_score, return_train_score, cache_cv)
    214     fields, tokens, params = normalize_params(candidate_params)
    215     main_token = tokenize(
--> 216         normalize_estimator(estimator),
    217         fields,
    218         params,

/raid/nicholasb/miniconda3/envs/rapids-21.10/lib/python3.8/site-packages/dask_ml/model_selection/_normalize.py in normalize_estimator(est)
     36                 continue
     37             try:
---> 38                 val = getattr(est, attr)
     39             except (sklearn.exceptions.NotFittedError, AttributeError):
     40                 continue

/raid/nicholasb/miniconda3/envs/rapids-21.10/lib/python3.8/site-packages/cuml/internals/api_decorators.py in inner(*args, **kwargs)
    593 
    594                 # Call the function
--> 595                 ret_val = func(*args, **kwargs)
    596 
    597             return cm.process_return(ret_val)

cuml/svm/svm_base.pyx in cuml.svm.svm_base.SVMBase.coef_()

RuntimeError: coef_ is only available for linear kernels

conda list | grep "rapids\|dask"
# packages in environment at /raid/nicholasb/miniconda3/envs/rapids-21.10:
cucim                     21.10.00a210907 cuda_11.2_py38_g27da55a_20    rapidsai-nightly
cudf                      21.10.00a210907 cuda_11.2_py38_g8d602b7e99_267    rapidsai-nightly
cudf_kafka                21.10.00a210907 py38_g8d602b7e99_267    rapidsai-nightly
cugraph                   21.10.00a210907 cuda11.2_py38_gdcc08bc6_67    rapidsai-nightly
cuml                      21.10.00a210907 cuda11.2_py38_g0e770fa50_71    rapidsai-nightly
cusignal                  21.10.00a210907 py38_g5ae8e6a_13    rapidsai-nightly
cuspatial                 21.10.00a210907 py38_gf11540e_13    rapidsai-nightly
custreamz                 21.10.00a210907 py38_g8d602b7e99_267    rapidsai-nightly
cuxfilter                 21.10.00a210907 py38_gd7204b9_13    rapidsai-nightly
dask                      2021.9.0           pyhd8ed1ab_0    conda-forge
dask-core                 2021.9.0           pyhd8ed1ab_0    conda-forge
dask-cuda                 21.10.00a210907         py38_35    rapidsai-nightly
dask-cudf                 21.10.00a210907 py38_g8d602b7e99_267    rapidsai-nightly
dask-glm                  0.2.0                      py_1    conda-forge
dask-ml                   1.9.0              pyhd8ed1ab_0    conda-forge
libcucim                  21.10.00a210907 cuda11.2_g27da55a_20    rapidsai-nightly
libcudf                   21.10.00a210907 cuda11.2_g8d602b7e99_267    rapidsai-nightly
libcudf_kafka             21.10.00a210907 g8d602b7e99_267    rapidsai-nightly
libcugraph                21.10.00a210907 cuda11.2_gdcc08bc6_67    rapidsai-nightly
libcuml                   21.10.00a210907 cuda11.2_g0e770fa50_71    rapidsai-nightly
libcumlprims              21.10.00a210826 cuda11.2_ga512fc5_5    rapidsai-nightly
libcuspatial              21.10.00a210907 cuda11.2_gf11540e_13    rapidsai-nightly
librmm                    21.10.00a210907 cuda11.2_g8527317_28    rapidsai-nightly
libxgboost                1.4.2dev.rapidsai21.10      cuda11.2_0    rapidsai-nightly
py-xgboost                1.4.2dev.rapidsai21.10  cuda11.2py38_0    rapidsai-nightly
rapids                    21.10.00a210907 cuda11.2_py38_g24de107_61    rapidsai-nightly
rapids-xgboost            21.10.00a210907 cuda11.2_py38_g24de107_61    rapidsai-nightly
rmm                       21.10.00a210907 cuda_11.2_py38_g8527317_28    rapidsai-nightly
ucx                       1.9.0+gcd9efd3       cuda11.2_0    rapidsai-nightly
ucx-proc                  1.0.0                       gpu    rapidsai-nightly
ucx-py                    0.22.0a210907   py38_gcd9efd3_18    rapidsai-nightly
xgboost                   1.4.2dev.rapidsai21.10  cuda11.2py38_0    rapidsai-nightly

[teju85]

Tagging @tfeher and @achirkin

[achirkin]

Thanks for the report! There seems to be nothing speaking against changing the error type here; I'm on it.

[beckernick]

Confirmed that the dask-ml code in the issue now works. cc @sauravdev