Enable multiclass svm for sparse input (#5588)

This commit enables multiclass SVM for sparse input. Previously this was deactivated as the `input_to_host_array` functionality does not support sparse arrays, but the data has to be piped through sklearn classes which requires host data. 

@dantegd , this is a local workaround to enable a `input_to_host_array` for sparse data without the complexity of providing the whole functionality of that function. Please have a look whether this is an acceptable solution for this use case.

FYI, @tfeher

Authors:
  - Malte Förster (https://github.com/mfoerste4)
  - Tamas Bela Feher (https://github.com/tfeher)
  - Simon Adorf (https://github.com/csadorf)

Approvers:
  - Tamas Bela Feher (https://github.com/tfeher)
  - Simon Adorf (https://github.com/csadorf)

URL: #5588

cpp/src/svm/kernelcache.cuh

@@ -130,8 +130,16 @@ class BatchCache : public raft::cache::Cache<math_t> {
     RAFT_CUDA_TRY(cudaMemsetAsync(tmp_buffer, 0, n_ws * 2 * sizeof(int), stream));
 
     // Init cub buffers
-    cub::DeviceRadixSort::SortKeys(
-      NULL, d_temp_storage_size, tmp_buffer, tmp_buffer, n_ws, 0, sizeof(int) * 8, stream);
+    cub::DeviceRadixSort::SortPairs(NULL,
+                                    d_temp_storage_size,
+                                    tmp_buffer,
+                                    tmp_buffer,
+                                    tmp_buffer,
+                                    tmp_buffer,
+                                    n_ws,
+                                    0,
+                                    sizeof(int) * 8,
+                                    stream);
     d_temp_storage.resize(d_temp_storage_size, stream);
   }
 

python/cuml/common/__init__.py

@@ -1,5 +1,5 @@
 #
-# Copyright (c) 2019-2022, NVIDIA CORPORATION.
+# Copyright (c) 2019-2023, NVIDIA CORPORATION.
 #
 # Licensed under the Apache License, Version 2.0 (the "License");
 # you may not use this file except in compliance with the License.
@@ -31,6 +31,7 @@
 
 from cuml.internals.input_utils import input_to_cuml_array
 from cuml.internals.input_utils import input_to_host_array
+from cuml.internals.input_utils import input_to_host_array_with_sparse_support
 
 from cuml.internals.memory_utils import rmm_cupy_ary
 from cuml.internals.memory_utils import set_global_output_type
@@ -59,6 +60,7 @@
     "has_scipy",
     "input_to_cuml_array",
     "input_to_host_array",
+    "input_to_host_array_with_sparse_support",
     "rmm_cupy_ary",
     "set_global_output_type",
     "using_device_type",

python/cuml/internals/input_utils.py

@@ -497,6 +497,20 @@ def input_to_host_array(
     return out_data._replace(array=out_data.array.to_output("numpy"))
 
 
+def input_to_host_array_with_sparse_support(X):
+    _array_type, is_sparse = determine_array_type_full(X)
+    if is_sparse:
+        if _array_type == "cupy":
+            return SparseCumlArray(X).to_output(output_type="scipy")
+        elif _array_type == "cuml":
+            return X.to_output(output_type="scipy")
+        elif _array_type == "numpy":
+            return X
+        else:
+            raise ValueError(f"Unsupported sparse array type: {_array_type}.")
+    return input_to_host_array(X).array
+
+
 def convert_dtype(X, to_dtype=np.float32, legacy=True, safe_dtype=True):
     """
     Convert X to be of dtype `dtype`, raising a TypeError

python/cuml/multiclass/multiclass.py

@@ -20,7 +20,15 @@
 from cuml.internals.import_utils import has_sklearn
 from cuml.internals.mixins import ClassifierMixin
 from cuml.common.doc_utils import generate_docstring
-from cuml.common import input_to_host_array
+from cuml.common import (
+    input_to_host_array,
+    input_to_host_array_with_sparse_support,
+)
+from cuml.internals.input_utils import (
+    input_to_cupy_array,
+    determine_array_type_full,
+)
+from cuml.internals.array_sparse import SparseCumlArray
 from cuml.internals import _deprecate_pos_args
 
 
@@ -142,7 +150,9 @@ def fit(self, X, y) -> "MulticlassClassifier":
                 + ", must be one of "
                 '{"ovr", "ovo"}'
             )
-        X = input_to_host_array(X).array
+
+        X = input_to_host_array_with_sparse_support(X)
+
         y = input_to_host_array(y).array
         with cuml.internals.exit_internal_api():
             self.multiclass_estimator.fit(X, y)
@@ -160,7 +170,8 @@ def predict(self, X) -> CumlArray:
         """
         Predict using multi class classifier.
         """
-        X = input_to_host_array(X).array
+        X = input_to_host_array_with_sparse_support(X)
+
         with cuml.internals.exit_internal_api():
             return self.multiclass_estimator.predict(X)
 
@@ -177,7 +188,7 @@ def decision_function(self, X) -> CumlArray:
         """
         Calculate the decision function.
         """
-        X = input_to_host_array(X).array
+        X = input_to_host_array_with_sparse_support(X)
         with cuml.internals.exit_internal_api():
             return self.multiclass_estimator.decision_function(X)
 

python/cuml/svm/svc.pyx

@@ -35,7 +35,7 @@ from cuml.common.doc_utils import generate_docstring
 from cuml.internals.logger import warn
 from pylibraft.common.handle cimport handle_t
 from pylibraft.common.interruptible import cuda_interruptible
-from cuml.common import input_to_cuml_array, input_to_host_array
+from cuml.common import input_to_cuml_array, input_to_host_array, input_to_host_array_with_sparse_support
 from cuml.internals.input_utils import input_to_cupy_array, determine_array_type_full
 from cuml.preprocessing import LabelEncoder
 from libcpp cimport nullptr
@@ -449,7 +449,7 @@ class SVC(SVMBase,
 
         # Currently CalibratedClassifierCV expects data on the host, see
         # https://github.com/rapidsai/cuml/issues/2608
-        X = input_to_host_array(X).array
+        X = input_to_host_array_with_sparse_support(X)
         y = input_to_host_array(y).array
 
         if not has_sklearn():
@@ -485,8 +485,6 @@ class SVC(SVMBase,
             return self._fit_proba(X, y, sample_weight)
 
         if self.n_classes_ > 2:
-            if is_sparse:
-                raise ValueError("Multiclass SVM does not support sparse input.")
             return self._fit_multiclass(X, y, sample_weight)
 
         if is_sparse:
@@ -594,7 +592,7 @@ class SVC(SVMBase,
         if self.probability:
             self._check_is_fitted('prob_svc')
 
-            X = input_to_host_array(X).array
+            X = input_to_host_array_with_sparse_support(X)
 
             with cuml.internals.exit_internal_api():
                 preds = self.prob_svc.predict(X)
@@ -628,7 +626,7 @@ class SVC(SVMBase,
         if self.probability:
             self._check_is_fitted('prob_svc')
 
-            X = input_to_host_array(X).array
+            X = input_to_host_array_with_sparse_support(X)
 
             # Exit the internal API when calling sklearn code (forces numpy
             # conversion)

python/cuml/tests/test_pickle.py

@@ -705,9 +705,6 @@ def assert_second_model(pickled_model, X):
 def test_svc_pickle(tmpdir, datatype, params, multiclass, sparse):
     result = {}
 
-    if sparse and multiclass:
-        pytest.skip("Multiclass SVC does not support sparse input")
-
     if sparse and params["probability"]:
         pytest.skip("Probabilistic SVC does not support sparse input")
 

python/cuml/tests/test_svm.py

@@ -41,8 +41,8 @@
 np = cpu_only_import("numpy")
 cuda = gpu_only_import_from("numba", "cuda")
 
-
 cudf = gpu_only_import("cudf")
+scipy_sparse = cpu_only_import("scipy.sparse")
 
 IS_ARM = platform.processor() == "aarch64"
 
@@ -176,13 +176,18 @@ def test_svm_skl_cmp_datasets(params, dataset, n_rows, n_cols):
 
 
 @pytest.mark.parametrize("params", [{"kernel": "rbf", "C": 1, "gamma": 1}])
+@pytest.mark.parametrize("sparse", [True, False])
 def test_svm_skl_cmp_multiclass(
-    params, dataset="classification2", n_rows=100, n_cols=6
+    params, sparse, dataset="classification2", n_rows=100, n_cols=6
 ):
     X_train, X_test, y_train, y_test = make_dataset(
         dataset, n_rows, n_cols, n_classes=3, n_informative=6
     )
 
+    if sparse:
+        X_train = scipy_sparse.csr_matrix(X_train)
+        X_test = scipy_sparse.csr_matrix(X_test)
+
     # Default to numpy for testing
     with cuml.using_output_type("numpy"):