Delay some sklearn imports

After this commit, `import cuml` should only import core bits of sklearn
like `import sklearn.base`. Anything algorithm specific is delayed until
first use, speeding import time.

Author: jcrist

python/cuml/cuml/explainer/kernel_shap.pyx

@@ -20,7 +20,6 @@ from random import randint
 
 import cupy as cp
 import numpy as np
-from sklearn.linear_model import LassoLarsIC, lars_path
 
 from cuml.explainer.base import SHAPBase
 from cuml.explainer.common import get_cai_ptr, model_func_call
@@ -565,6 +564,7 @@ def _l1_regularization(X,
     """
     Function calls LASSO or LARS if l1 regularization is needed.
     """
+    from sklearn.linear_model import LassoLarsIC, lars_path
 
     # create augmented dataset for feature selection
     s = cp.sum(X, axis=1)

python/cuml/cuml/explainer/tree_shap.pyx

@@ -14,13 +14,13 @@
 # limitations under the License.
 #
 
+import cuml
 from cuml.common import input_to_cuml_array
-from cuml.ensemble import RandomForestClassifier as curfc
-from cuml.ensemble import RandomForestRegressor as curfr
 from cuml.internals.array import CumlArray
 from cuml.internals.input_utils import determine_array_type
 
 from cuml.internals.treelite cimport *
+
 from cuml.internals.treelite import safe_treelite_call
 
 from libc.stdint cimport uintptr_t
@@ -29,8 +29,9 @@ import re
 
 import numpy as np
 import treelite
-from sklearn.ensemble import RandomForestClassifier as sklrfc
-from sklearn.ensemble import RandomForestRegressor as sklrfr
+
+# from sklearn.ensemble import RandomForestClassifier as sklrfc
+# from sklearn.ensemble import RandomForestRegressor as sklrfr
 
 
 cdef extern from "cuml/explainer/tree_shap.hpp" namespace "ML::Explainer" nogil:
@@ -178,15 +179,20 @@ cdef class TreeExplainer:
                 model = model.booster_
             tl_model = treelite.frontend.from_lightgbm(model)
         # cuML RF model object
-        elif isinstance(model, (curfr, curfc)):
+        elif isinstance(model, (cuml.RandomForestClassifier, cuml.RandomForestRegressor)):
             tl_model = model.convert_to_treelite_model()
         # scikit-learn RF model object
-        elif isinstance(model, (sklrfr, sklrfc)):
-            tl_model = treelite.sklearn.import_model(model)
         elif isinstance(model, treelite.Model):
             tl_model = model
         else:
-            raise ValueError(f"Unrecognized model object type: {type(model)}")
+            from sklearn.ensemble import (
+                RandomForestClassifier,
+                RandomForestRegressor,
+            )
+            if isinstance(model, (RandomForestClassifier, RandomForestRegressor)):
+                tl_model = treelite.sklearn.import_model(model)
+            else:
+                raise ValueError(f"Unrecognized model object type: {type(model)}")
 
         # Get num_class
         self.num_class = tl_model.get_header_accessor().get_field("num_class").copy()

python/cuml/cuml/internals/base.py

@@ -18,7 +18,6 @@
 import os
 
 import pylibraft.common.handle
-from sklearn.utils import estimator_html_repr
 
 import cuml
 import cuml.common
@@ -458,10 +457,11 @@ def _more_tags(self):
 
     def _repr_mimebundle_(self, **kwargs):
         """Prepare representations used by jupyter kernels to display estimator"""
-        if estimator_html_repr is not None:
-            output = {"text/plain": repr(self)}
-            output["text/html"] = estimator_html_repr(self)
-            return output
+        from sklearn.utils import estimator_html_repr
+
+        output = {"text/plain": repr(self)}
+        output["text/html"] = estimator_html_repr(self)
+        return output
 
     def set_nvtx_annotations(self):
         for func_name in [

python/cuml/cuml/random_projection/random_projection.py

@@ -15,7 +15,6 @@
 import cupyx.scipy.sparse as cp_sp
 import numpy as np
 import scipy.sparse as sp
-import sklearn.random_projection
 
 import cuml
 from cuml.common.array_descriptor import CumlArrayDescriptor
@@ -51,6 +50,8 @@ def johnson_lindenstrauss_min_dim(n_samples, eps=0.1):
         The minimal number of components to guarantee with good probability
         an eps-embedding with n_samples.
     """
+    import sklearn.random_projection
+
     return sklearn.random_projection.johnson_lindenstrauss_min_dim(
         n_samples, eps=eps
     )