KNN SPMD python interfaces (#1208)

* initial draft of python knn SPMD interfaces * conditionally adding responses to regression * forgot this * trying moving get_queue * reverting last and restoring cpp to test * trying just infer * resolved issue and examples run * removing commented lines * ci update * reverting intel-dpcpp-cpp-compiler version change * reverting * reverting several commits... * addressing comments * flake8 * addressing comments * flake8 * specifying onedal version for spmd neighbors setup * temporary commit for debugging external CI fails * printing import error * adding conditional to neighbors.cp * adding version import * removing all logging and re-adding onedal version conditional in setup * addressing last comments * removing comm size warnings and cleaning up examples * example rename * update run_examples.py after rebase * flake8 * addressing non gpu spmd call, better multiline string
uxlfoundation · Mar 23, 2023 · 6ef0a5d · 6ef0a5d
1 parent cfc214c
commit 6ef0a5d
Show file tree

Hide file tree

Showing 13 changed files with 283 additions and 10 deletions.
diff --git a/examples/sklearnex/knn_bf_classification_spmd.py b/examples/sklearnex/knn_bf_classification_spmd.py
@@ -0,0 +1,65 @@
+#===============================================================================
+# Copyright 2023 Intel Corporation
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+#===============================================================================
+
+import numpy as np
+from sklearn.metrics import accuracy_score
+from warnings import warn
+from mpi4py import MPI
+import dpctl
+from sklearnex.spmd.neighbors import KNeighborsClassifier
+
+
+def generate_X_y(par, seed):
+    ns, nf = par['ns'], par['nf']
+
+    drng = np.random.default_rng(seed)
+    data = drng.uniform(-1, 1, size=(ns, nf))
+    resp = (data > 0) @ (2 ** np.arange(nf))
+
+    return data, resp
+
+
+comm = MPI.COMM_WORLD
+rank = comm.Get_rank()
+size = comm.Get_size()
+
+if dpctl.has_gpu_devices:
+    q = dpctl.SyclQueue("gpu")
+else:
+    raise RuntimeError('GPU devices unavailable. Currently, '
+          'SPMD execution mode is implemented only for this device type.')
+
+params_train = {'ns': 100000, 'nf': 8}
+params_test = {'ns': 100, 'nf': 8}
+
+X_train, y_train = generate_X_y(params_train, rank)
+X_test, y_test = generate_X_y(params_test, rank + 99)
+
+model_spmd = KNeighborsClassifier(algorithm='brute',
+                                  n_neighbors=20,
+                                  weights='uniform',
+                                  p=2,
+                                  metric='minkowski')
+model_spmd.fit(X_train, y_train, queue=q)
+
+y_predict = model_spmd.predict(X_test, queue=q)
+
+print("Brute Force Distributed kNN classification results:")
+print("Ground truth (first 5 observations on rank {}):\n{}".format(rank, y_test[:5]))
+print("Classification results (first 5 observations on rank {}):\n{}"
+      .format(rank, y_predict[:5]))
+print("Accuracy for entire rank {} (256 classes): {}\n"
+      .format(rank, accuracy_score(y_test, y_predict)))
diff --git a/examples/sklearnex/knn_bf_regression_spmd.py b/examples/sklearnex/knn_bf_regression_spmd.py
@@ -0,0 +1,71 @@
+#===============================================================================
+# Copyright 2023 Intel Corporation
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+#===============================================================================
+
+import numpy as np
+from sklearn.metrics import mean_squared_error
+from warnings import warn
+from mpi4py import MPI
+import dpctl
+from numpy.testing import assert_allclose
+from sklearnex.spmd.neighbors import KNeighborsRegressor
+
+
+def generate_X_y(par, coef_seed, data_seed):
+    ns, nf = par['ns'], par['nf']
+
+    crng = np.random.default_rng(coef_seed)
+    coef = crng.uniform(-10, 10, size=(nf,))
+
+    drng = np.random.default_rng(data_seed)
+    data = drng.uniform(-100, 100, size=(ns, nf))
+    resp = data @ coef
+
+    return data, resp, coef
+
+
+comm = MPI.COMM_WORLD
+rank = comm.Get_rank()
+size = comm.Get_size()
+
+if dpctl.has_gpu_devices:
+    q = dpctl.SyclQueue("gpu")
+else:
+    raise RuntimeError('GPU devices unavailable. Currently, '
+          'SPMD execution mode is implemented only for this device type.')
+
+params_train = {'ns': 1000000, 'nf': 3}
+params_test = {'ns': 100, 'nf': 3}
+
+X_train, y_train, coef_train = generate_X_y(params_train, 10, rank)
+X_test, y_test, coef_test = generate_X_y(params_test, 10, rank + 99)
+
+assert_allclose(coef_train, coef_test)
+
+model_spmd = KNeighborsRegressor(algorithm='brute',
+                                 n_neighbors=5,
+                                 weights='uniform',
+                                 p=2,
+                                 metric='minkowski')
+model_spmd.fit(X_train, y_train, queue=q)
+
+y_predict = model_spmd.predict(X_test, queue=q)
+
+print("Brute Force Distributed kNN regression results:")
+print("Ground truth (first 5 observations on rank {}):\n{}".format(rank, y_test[:5]))
+print("Regression results (first 5 observations on rank {}):\n{}"
+      .format(rank, y_predict[:5]))
+print("RMSE for entire rank {}: {}\n"
+      .format(rank, mean_squared_error(y_test, y_predict, squared=False)))
diff --git a/onedal/__init__.py b/onedal/__init__.py
@@ -49,4 +49,5 @@
     __all__ += ['basic_statistics', 'linear_model']
 
     if _is_dpc_backend:
-        __all__ += ['spmd.basic_statistics', 'spmd.decomposition', 'spmd.linear_model',]
+        __all__ += ['spmd.basic_statistics', 'spmd.decomposition',
+                    'spmd.linear_model', 'spmd.neighbors']
diff --git a/onedal/neighbors/neighbors.cpp b/onedal/neighbors/neighbors.cpp
@@ -17,6 +17,7 @@
 #include "oneapi/dal/algo/knn.hpp"
 
 #include "onedal/common.hpp"
+#include "onedal/version.hpp"
 #include "onedal/primitives/pairwise_distances.hpp"
 #include <regex>
 
@@ -313,8 +314,13 @@ ONEDAL_PY_INIT_MODULE(neighbors) {
     using task_list = types<task::classification, task::regression, task::search>;
     auto sub = m.def_submodule("neighbors");
 
+#if defined(ONEDAL_DATA_PARALLEL_SPMD) && defined(ONEDAL_VERSION) && ONEDAL_VERSION >= 20230100
+    ONEDAL_PY_INSTANTIATE(init_train_ops, sub, policy_list_spmd, task_list);
+    ONEDAL_PY_INSTANTIATE(init_infer_ops, sub, policy_list_spmd, task_list);
+#else // defined(ONEDAL_DATA_PARALLEL_SPMD) && defined(ONEDAL_VERSION) && ONEDAL_VERSION >= 20230100
     ONEDAL_PY_INSTANTIATE(init_train_ops, sub, policy_list, task_list);
     ONEDAL_PY_INSTANTIATE(init_infer_ops, sub, policy_list, task_list);
+#endif // defined(ONEDAL_DATA_PARALLEL_SPMD) && defined(ONEDAL_VERSION) && ONEDAL_VERSION >= 20230100
 
     ONEDAL_PY_INSTANTIATE(init_model, sub, task_list);
     ONEDAL_PY_INSTANTIATE(init_train_result, sub, task_list);

diff --git a/onedal/neighbors/neighbors.py b/onedal/neighbors/neighbors.py
@@ -44,6 +44,9 @@
 
 
 class NeighborsCommonBase(metaclass=ABCMeta):
+    def _get_policy(self, queue, *data):
+        return _get_policy(queue, *data)
+
     def _parse_auto_method(self, method, n_samples, n_features):
         result_method = method
 
@@ -402,7 +405,7 @@ def _onedal_fit(self, X, y, queue):
 
             return train_alg(**params).compute(X, y).model
 
-        policy = _get_policy(queue, X, y)
+        policy = self._get_policy(queue, X, y)
         X, y = _convert_to_supported(policy, X, y)
         params = self._get_onedal_params(X, y)
         train_alg = _backend.neighbors.classification.train(policy, params,
@@ -421,7 +424,7 @@ def _onedal_predict(self, model, X, params, queue):
 
             return predict_alg(**params).compute(X, model)
 
-        policy = _get_policy(queue, X)
+        policy = self._get_policy(queue, X)
         X = _convert_to_supported(policy, X)
         if hasattr(self, '_onedal_model'):
             model = self._onedal_model
@@ -549,7 +552,8 @@ def _onedal_fit(self, X, y, queue):
 
             return train_alg(**params).compute(X, y).model
 
-        policy = _get_policy(queue, X, y)
+        policy = self._get_policy(queue, X, y)
+        X, y = _convert_to_supported(policy, X, y)
         params = self._get_onedal_params(X, y)
         train_alg_regr = _backend.neighbors.regression.train
         train_alg_srch = _backend.neighbors.search.train
@@ -568,7 +572,8 @@ def _onedal_predict(self, model, X, params, queue):
 
             return predict_alg(**params).compute(X, model)
 
-        policy = _get_policy(queue, X)
+        policy = self._get_policy(queue, X)
+        X = _convert_to_supported(policy, X)
         backend = _backend.neighbors.regression if gpu_device \
             else _backend.neighbors.search
 
@@ -678,7 +683,8 @@ def _onedal_fit(self, X, y, queue):
 
             return train_alg(**params).compute(X, y).model
 
-        policy = _get_policy(queue, X, y)
+        policy = self._get_policy(queue, X, y)
+        X, y = _convert_to_supported(policy, X, y)
         params = self._get_onedal_params(X, y)
         train_alg = _backend.neighbors.search.train(policy, params,
                                                     to_table(X))
@@ -696,7 +702,8 @@ def _onedal_predict(self, model, X, params, queue):
 
             return predict_alg(**params).compute(X, model)
 
-        policy = _get_policy(queue, X)
+        policy = self._get_policy(queue, X)
+        X = _convert_to_supported(policy, X)
         if hasattr(self, '_onedal_model'):
             model = self._onedal_model
         else:

diff --git a/onedal/spmd/__init__.py b/onedal/spmd/__init__.py
@@ -14,4 +14,4 @@
 # limitations under the License.
 #===============================================================================
 
-__all__ = ['basic_statistics', 'decomposition', 'linear_model']
+__all__ = ['basic_statistics', 'decomposition', 'linear_model', 'neighbors']
diff --git a/onedal/spmd/neighbors/__init__.py b/onedal/spmd/neighbors/__init__.py
@@ -0,0 +1,19 @@
+#===============================================================================
+# Copyright 2023 Intel Corporation
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+#===============================================================================
+
+from .neighbors import KNeighborsClassifier, KNeighborsRegressor, NearestNeighbors
+
+__all__ = ['KNeighborsClassifier', 'KNeighborsRegressor', 'NearestNeighbors']
diff --git a/onedal/spmd/neighbors/neighbors.py b/onedal/spmd/neighbors/neighbors.py
@@ -0,0 +1,52 @@
+#===============================================================================
+# Copyright 2023 Intel Corporation
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+#===============================================================================
+
+from abc import ABC
+from ...common._spmd_policy import _get_spmd_policy
+from onedal.neighbors import KNeighborsClassifier as KNeighborsClassifier_Batch
+from onedal.neighbors import KNeighborsRegressor as KNeighborsRegressor_Batch
+
+
+class NeighborsCommonBaseSPMD(ABC):
+    def _get_policy(self, queue, *data):
+        return _get_spmd_policy(queue)
+
+
+class KNeighborsClassifier(NeighborsCommonBaseSPMD, KNeighborsClassifier_Batch):
+    def predict_proba(self, X, queue=None):
+        raise NotImplementedError("predict_proba not supported in distributed mode.")
+
+
+class KNeighborsRegressor(NeighborsCommonBaseSPMD, KNeighborsRegressor_Batch):
+    def fit(self, X, y, queue=None):
+        if queue is not None and queue.sycl_device.is_gpu:
+            return super()._fit(X, y, queue=queue)
+        else:
+            raise ValueError('SPMD version of kNN is not implemented for '
+                  'CPU. Consider running on it on GPU.')
+
+    def predict(self, X, queue=None):
+        return self._predict_gpu(X, queue=queue)
+
+    def _get_onedal_params(self, X, y=None):
+        params = super()._get_onedal_params(X, y)
+        if 'responses' not in params['result_option']:
+            params['result_option'] += '|responses'
+        return params
+
+
+class NearestNeighbors(NeighborsCommonBaseSPMD):
+    pass
diff --git a/setup.py b/setup.py
@@ -478,7 +478,9 @@ def run(self):
          'onedal.spmd.basic_statistics',
          'onedal.spmd.decomposition',
          'onedal.spmd.linear_model'
-         ] if build_distribute else [])),
+         ] + (['onedal.spmd.neighbors']
+              if ONEDAL_VERSION >= 20230100 else [])
+        if build_distribute else [])),
     package_data={
         'daal4py.oneapi': [
             'liboneapi_backend.so',

diff --git a/sklearnex/spmd/__init__.py b/sklearnex/spmd/__init__.py
@@ -14,4 +14,4 @@
 # limitations under the License.
 #===============================================================================
 
-__all__ = ['basic_statistics', 'decomposition', 'linear_model']
+__all__ = ['basic_statistics', 'decomposition', 'linear_model', 'neighbors']
diff --git a/sklearnex/spmd/neighbors/__init__.py b/sklearnex/spmd/neighbors/__init__.py
@@ -0,0 +1,19 @@
+#===============================================================================
+# Copyright 2023 Intel Corporation
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+#===============================================================================
+
+from .neighbors import KNeighborsClassifier, KNeighborsRegressor, NearestNeighbors
+
+__all__ = ['KNeighborsClassifier', 'KNeighborsRegressor', 'NearestNeighbors']
diff --git a/sklearnex/spmd/neighbors/neighbors.py b/sklearnex/spmd/neighbors/neighbors.py
@@ -0,0 +1,25 @@
+#===============================================================================
+# Copyright 2023 Intel Corporation
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+#===============================================================================
+
+from onedal.spmd.neighbors import (
+    KNeighborsClassifier,
+    KNeighborsRegressor,
+    NearestNeighbors
+)
+
+# TODO:
+# Currently it uses `onedal` module interface.
+# Add sklearnex dispatching.