[Hackathon No.18] 为 Paddle 新增 frexp API (#46401)

* 之前的pr合并了大量错误代码，重新提交一份

* 之前的pr合并了大量错误代码，重新提交一份

* 修正格式问题

* 改回原来的格式

* 按照要求修改

* 按照要求修改格式

* 修复注释的问题

* 更新格式

* 测试自动格式化

* 修正英文注释

* fix docs build error

* pre-commit

* for docs build

* for docs build

* 修复mantissa计算错误的bug

* 修复误判exponent可能存在负数，导致计算量增加的情况

Co-authored-by: Ligoml <[email protected]>

Author: Zheng-Bicheng

python/paddle/__init__.py

@@ -286,6 +286,7 @@
 from .tensor.math import frac  # noqa: F401
 from .tensor.math import sgn  # noqa: F401
 from .tensor.math import take  # noqa: F401
+from .tensor.math import frexp  # noqa: F401
 
 from .tensor.random import bernoulli  # noqa: F401
 from .tensor.random import poisson  # noqa: F401
@@ -386,7 +387,6 @@
         os.environ.setdefault('runtime_include_dir', runtime_include_dir)
 
 disable_static()
-
 __all__ = [  # noqa
     'iinfo',
     'dtype',
@@ -667,4 +667,5 @@
     'sgn',
     'triu_indices',
     'take',
+    'frexp',
 ]

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

@@ -0,0 +1,94 @@
+# 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.
+import unittest
+import numpy as np
+import paddle
+import paddle.fluid
+
+
+class TestFrexpAPI(unittest.TestCase):
+
+    def setUp(self):
+        np.random.seed(1024)
+        self.rtol = 1e-5
+        self.atol = 1e-8
+        self.place = paddle.CUDAPlace(0) if paddle.is_compiled_with_cuda() \
+            else paddle.CPUPlace()
+        self.set_input()
+
+    def set_input(self):
+        self.x_np = np.random.uniform(-3, 3, [10, 12]).astype('float32')
+
+    # 静态图单测
+    def test_static_api(self):
+        # 开启静态图模式
+        paddle.enable_static()
+        with paddle.static.program_guard(paddle.static.Program()):
+            input_data = paddle.fluid.data('X', self.x_np.shape,
+                                           self.x_np.dtype)
+            out = paddle.frexp(input_data)
+            # 计算静态图结果
+            exe = paddle.static.Executor(self.place)
+            res = exe.run(feed={'X': self.x_np}, fetch_list=[out])
+
+        out_ref = np.frexp(self.x_np)
+        # 对比静态图与 numpy 实现函数计算结果是否相同
+        for n, p in zip(out_ref, res):
+            np.testing.assert_allclose(n, p, rtol=self.rtol, atol=self.atol)
+
+    # 动态图单测
+    def test_dygraph_api(self):
+        # 关闭静态图模式
+        paddle.disable_static(self.place)
+        input_num = paddle.to_tensor(self.x_np)
+        # 测试动态图 tensor.frexp 和 paddle.tensor.math.frexp 计算结果
+        out1 = np.frexp(self.x_np)
+        out2 = paddle.frexp(input_num)
+        np.testing.assert_allclose(out1, out2, rtol=1e-05)
+
+        out1 = np.frexp(self.x_np)
+        out2 = input_num.frexp()
+        np.testing.assert_allclose(out1, out2, rtol=1e-05)
+        paddle.enable_static()
+
+
+class TestSplitsFloat32Case1(TestFrexpAPI):
+    """
+        Test num_or_sections which is an integer and data type is float32.
+    """
+
+    def set_input(self):
+        self.x_np = np.random.uniform(-1, 1, [4, 5, 2]).astype('float32')
+
+
+class TestSplitsFloat64Case1(TestFrexpAPI):
+    """
+        Test num_or_sections which is an integer and data type is float64.
+    """
+
+    def set_input(self):
+        self.x_np = np.random.uniform(-3, 3, [10, 12]).astype('float64')
+
+
+class TestSplitsFloat64Case2(TestFrexpAPI):
+    """
+        Test num_or_sections which is an integer and data type is float64.
+    """
+
+    def set_input(self):
+        self.x_np = np.random.uniform(-1, 1, [4, 5, 2]).astype('float64')
+
+
+if __name__ == "__main__":
+    unittest.main()

python/paddle/tensor/__init__.py

@@ -239,6 +239,7 @@
 from .math import frac  # noqa: F401
 from .math import sgn  # noqa: F401
 from .math import take  # noqa: F401
+from .math import frexp  # noqa: F401
 
 from .random import multinomial  # noqa: F401
 from .random import standard_normal  # noqa: F401
@@ -517,6 +518,7 @@
     'take',
     'bucketize',
     'sgn',
+    'frexp',
 ]
 
 # this list used in math_op_patch.py for magic_method bind

python/paddle/tensor/math.py

@@ -5108,3 +5108,52 @@ def take(x, index, mode='raise', name=None):
     out = input_1d.index_select(index_1d).reshape(index.shape)
 
     return out
+
+
+def frexp(x, name=None):
+    """
+    The function used to decompose a floating point number into mantissa and exponent.
+
+    Args:
+        x (Tensor): The input tensor, it's data type should be float32, float64.
+        name (str, optional): For details, please refer to :ref:`api_guide_Name`. Generally, no setting is required. Default: None.
+    Returns:
+
+        - mantissa (Tensor), A mantissa Tensor. The shape and data type of mantissa tensor and exponential tensor are
+            the same as those of input.
+
+        - exponent (Tensor), A exponent Tensor. The shape and data type of mantissa tensor and exponential tensor are
+            the same as those of input.
+
+    Examples:
+        .. code-block:: python
+
+            import paddle
+
+            x = paddle.to_tensor([[1, 2, 3, 4]], dtype="float32")
+            print(paddle.tensor.math.frexp(x))
+            # (Tensor(shape=[1, 4], dtype=float32, place=Place(cpu), stop_gradient=True,[[0.50000000, 0.50000000, 0.75000000, 0.50000000]]),
+            #  Tensor(shape=[1, 4], dtype=float32, place=Place(cpu), stop_gradient=True,[[1., 2., 2., 3.]]))
+        """
+    if x.dtype not in [paddle.float32, paddle.float64]:
+        raise TypeError(
+            "The data type of input must be one of ['float32', 'float64'], but got {}"
+            .format(x.dtype))
+    input_x = paddle.abs(x)
+    exponent = paddle.floor(paddle.log2(input_x))
+    exponent = paddle.where(paddle.isinf(exponent),
+                            paddle.full_like(exponent, 0), exponent)
+
+    # 0填充
+    mantissa = paddle.divide(input_x, 2**exponent)
+    # 计算exponent
+    exponent = paddle.where((mantissa >= 1),
+                            paddle.add(exponent, paddle.ones_like(exponent)),
+                            exponent)
+    mantissa = paddle.where((mantissa >= 1),
+                            paddle.divide(mantissa,
+                                          2**paddle.ones_like(exponent)),
+                            mantissa)
+
+    mantissa = paddle.where((x < 0), mantissa * -1, mantissa)
+    return mantissa, exponent