【PaddlePaddle Hackathon 3 No.5】为 Paddle 新增 bucketize (#44195)

* add paddle.bucketize api

add paddle.bucketize api

* updata paddle.bucketize code style.

* upgrade unittests/test_bucketize_api.py.

* update tests/unittests/test_bucketize_api.py file.

* 修改bucketize API 的英文参数描述

python/paddle/__init__.py

@@ -292,6 +292,7 @@ from .tensor.search import argmax  # noqa: F401
 from .tensor.search import argmin  # noqa: F401
 from .tensor.search import argsort  # noqa: F401
 from .tensor.search import searchsorted  # noqa: F401
+from .tensor.search import bucketize  # noqa: F401
 from .tensor.search import masked_select  # noqa: F401
 from .tensor.search import topk  # noqa: F401
 from .tensor.search import where  # noqa: F401
@@ -443,6 +444,7 @@ __all__ = [  # noqa
     'flops',
     'sort',
     'searchsorted',
+    'bucketize',
     'split',
     'logical_and',
     'full_like',

python/paddle/fluid/tests/unittests/test_bucketize_api.py

+#   Copyright (c) 2022 PaddlePaddle Authors. All Rights Reserved.
+#
+# 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 __future__ import print_function
+from re import X
+
+import unittest
+import numpy as np
+import paddle
+import paddle.fluid as fluid
+import paddle.fluid.core as core
+from paddle.fluid import Program, program_guard
+
+np.random.seed(10)
+
+
+class TestBucketizeAPI(unittest.TestCase):
+    # test paddle.tensor.math.nanmean
+
+    def setUp(self):
+        self.sorted_sequence = np.array([2, 4, 8, 16]).astype("float64")
+        self.x = np.array([[0, 8, 4, 16], [-1, 2, 8, 4]]).astype("float64")
+        self.place = [paddle.CPUPlace()]
+        if core.is_compiled_with_cuda():
+            self.place.append(paddle.CUDAPlace(0))
+
+    def test_api_static(self):
+        paddle.enable_static()
+
+        def run(place):
+            with paddle.static.program_guard(paddle.static.Program()):
+                sorted_sequence = paddle.static.data(
+                    'SortedSequence',
+                    shape=self.sorted_sequence.shape,
+                    dtype="float64")
+                x = paddle.static.data('x', shape=self.x.shape, dtype="float64")
+                out1 = paddle.bucketize(x, sorted_sequence)
+                out2 = paddle.bucketize(x, sorted_sequence, right=True)
+                exe = paddle.static.Executor(place)
+                res = exe.run(feed={
+                    'SortedSequence': self.sorted_sequence,
+                    'x': self.x
+                },
+                              fetch_list=[out1, out2])
+            out_ref = np.searchsorted(self.sorted_sequence, self.x)
+            out_ref1 = np.searchsorted(self.sorted_sequence,
+                                       self.x,
+                                       side='right')
+            self.assertTrue(np.allclose(out_ref, res[0]))
+            self.assertTrue(np.allclose(out_ref1, res[1]))
+
+        for place in self.place:
+            run(place)
+
+    def test_api_dygraph(self):
+
+        def run(place):
+            paddle.disable_static(place)
+            sorted_sequence = paddle.to_tensor(self.sorted_sequence)
+            x = paddle.to_tensor(self.x)
+            out1 = paddle.bucketize(x, sorted_sequence)
+            out2 = paddle.bucketize(x, sorted_sequence, right=True)
+            out_ref1 = np.searchsorted(self.sorted_sequence, self.x)
+            out_ref2 = np.searchsorted(self.sorted_sequence,
+                                       self.x,
+                                       side='right')
+            self.assertEqual(np.allclose(out_ref1, out1.numpy()), True)
+            self.assertEqual(np.allclose(out_ref2, out2.numpy()), True)
+            paddle.enable_static()
+
+        for place in self.place:
+            run(place)
+
+    def test_out_int32(self):
+        paddle.disable_static()
+        sorted_sequence = paddle.to_tensor(self.sorted_sequence)
+        x = paddle.to_tensor(self.x)
+        out = paddle.bucketize(x, sorted_sequence, out_int32=True)
+        self.assertTrue(out.type, 'int32')
+
+    def test_bucketize_dims_error(self):
+        with paddle.static.program_guard(paddle.static.Program()):
+            sorted_sequence = paddle.static.data('SortedSequence',
+                                                 shape=[2, 2],
+                                                 dtype="float64")
+            x = paddle.static.data('x', shape=[2, 5], dtype="float64")
+            self.assertRaises(ValueError, paddle.bucketize, x, sorted_sequence)
+
+    def test_input_error(self):
+        for place in self.place:
+            paddle.disable_static(place)
+            sorted_sequence = paddle.to_tensor(self.sorted_sequence)
+            self.assertRaises(ValueError, paddle.bucketize, self.x,
+                              sorted_sequence)
+
+    def test_empty_input_error(self):
+        for place in self.place:
+            paddle.disable_static(place)
+            sorted_sequence = paddle.to_tensor(self.sorted_sequence)
+            x = paddle.to_tensor(self.x)
+            self.assertRaises(ValueError, paddle.bucketize, None,
+                              sorted_sequence)
+            self.assertRaises(AttributeError, paddle.bucketize, x, None)
+
+
+if __name__ == "__main__":
+    unittest.main()

python/paddle/tensor/__init__.py

@@ -250,6 +250,7 @@ from .search import argmax  # noqa: F401
 from .search import argmin  # noqa: F401
 from .search import argsort  # noqa: F401
 from .search import searchsorted  # noqa: F401
+from .search import bucketize  # noqa: F401
 from .search import topk  # noqa: F401
 from .search import where  # noqa: F401
 from .search import index_select  # noqa: F401
@@ -503,6 +504,7 @@ tensor_method_func = [  #noqa
     'put_along_axis_',
     'exponential_',
     'heaviside',
+    'bucketize',
 ]
 
 #this list used in math_op_patch.py for magic_method bind

python/paddle/tensor/search.py

@@ -914,6 +914,61 @@ def topk(x, k, axis=None, largest=True, sorted=True, name=None):
     return values, indices
 
 
+def bucketize(x, sorted_sequence, out_int32=False, right=False, name=None):
+    """
+    This API is used to find the index of the corresponding 1D tensor `sorted_sequence` in the innermost dimension based on the given `x`.
+
+    Args:
+        x(Tensor): An input N-D tensor value with type int32, int64, float32, float64.
+        sorted_sequence(Tensor): An input 1-D tensor with type int32, int64, float32, float64. The value of the tensor monotonically increases in the innermost dimension. 
+        out_int32(bool, optional): Data type of the output tensor which can be int32, int64. The default value is False, and it indicates that the output data type is int64.
+        right(bool, optional): Find the upper or lower bounds of the sorted_sequence range in the innermost dimension based on the given `x`. If the value of the sorted_sequence is nan or inf, return the size of the innermost dimension.
+                               The default value is False and it shows the lower bounds.  
+        name(str, optional): The default value is None. Normally there is no need for user to set this property. For more information, please refer to :ref:`api_guide_Name`.
+        
+    Returns:
+        Tensor（the same sizes of the `x`）, return the tensor of int32 if set :attr:`out_int32` is True, otherwise return the tensor of int64.  
+    
+    Examples:
+
+        .. code-block:: python
+    
+            import paddle
+
+            sorted_sequence = paddle.to_tensor([2, 4, 8, 16], dtype='int32')
+            x = paddle.to_tensor([[0, 8, 4, 16], [-1, 2, 8, 4]], dtype='int32')
+            out1 = paddle.bucketize(x, sorted_sequence)
+            print(out1)
+            # Tensor(shape=[2, 4], dtype=int64, place=CPUPlace, stop_gradient=True,
+            #        [[0, 2, 1, 3],
+            #         [0, 0, 2, 1]])
+            out2 = paddle.bucketize(x, sorted_sequence, right=True)
+            print(out2)
+            # Tensor(shape=[2, 4], dtype=int64, place=CPUPlace, stop_gradient=True,
+            #        [[0, 3, 2, 4],
+            #         [0, 1, 3, 2]])
+            out3 = x.bucketize(sorted_sequence)
+            print(out3)
+            # Tensor(shape=[2, 4], dtype=int64, place=CPUPlace, stop_gradient=True,
+            #        [[0, 2, 1, 3],
+            #         [0, 0, 2, 1]])
+            out4 = x.bucketize(sorted_sequence, right=True)
+            print(out4)
+            # Tensor(shape=[2, 4], dtype=int64, place=CPUPlace, stop_gradient=True,
+            #        [[0, 3, 2, 4],
+            #         [0, 1, 3, 2]])
+            
+    """
+    check_variable_and_dtype(sorted_sequence, 'SortedSequence',
+                             ['float32', 'float64', 'int32', 'int64'],
+                             'paddle.searchsorted')
+    if sorted_sequence.dim() != 1:
+        raise ValueError(
+            f"sorted_sequence tensor must be 1 dimension, but got dim {sorted_sequence.dim()}"
+        )
+    return searchsorted(sorted_sequence, x, out_int32, right, name)
+
+
 def searchsorted(sorted_sequence,
                  values,
                  out_int32=False,