【Hackathon No.6】implement nan_to_num (#42469)

python/paddle/__init__.py

@@ -230,6 +230,7 @@
 from .tensor.math import square  # noqa: F401
 from .tensor.math import stanh  # noqa: F401
 from .tensor.math import sum  # noqa: F401
+from .tensor.math import nan_to_num  # noqa: F401
 from .tensor.math import nansum  # noqa: F401
 from .tensor.math import nanmean  # noqa: F401
 from .tensor.math import count_nonzero  # noqa: F401
@@ -666,6 +667,7 @@
     'renorm',
     'take_along_axis',
     'put_along_axis',
+    'nan_to_num',
     'heaviside',
     'tril_indices',
     'index_add',

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

@@ -0,0 +1,203 @@
+#   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
+from typing import Optional
+import numpy as np
+import paddle
+import paddle.fluid.core as core
+
+# from op_test import OpTest
+
+
+def np_nan_to_num(
+    x: np.ndarray,
+    nan: float = 0.0,
+    posinf: Optional[float] = None,
+    neginf: Optional[float] = None,
+) -> np.ndarray:
+    return np.nan_to_num(x, True, nan=nan, posinf=posinf, neginf=neginf)
+
+
+def np_nan_to_num_op(
+    x: np.ndarray,
+    nan: float,
+    replace_posinf_with_max: bool,
+    posinf: float,
+    replace_neginf_with_min: bool,
+    neginf: float,
+) -> np.ndarray:
+    if replace_posinf_with_max:
+        posinf = None
+    if replace_neginf_with_min:
+        neginf = None
+    return np.nan_to_num(x, True, nan=nan, posinf=posinf, neginf=neginf)
+
+
+def np_nan_to_num_grad(x: np.ndarray, dout: np.ndarray) -> np.ndarray:
+    dx = np.copy(dout)
+    dx[np.isnan(x) | (x == np.inf) | (x == -np.inf)] = 0
+    return dx
+
+
+class TestNanToNum(unittest.TestCase):
+    def setUp(self):
+        self.place = (
+            paddle.CUDAPlace(0)
+            if core.is_compiled_with_cuda()
+            else paddle.CPUPlace()
+        )
+
+    def test_static(self):
+        x_np = np.array([[1, np.nan, -2], [np.inf, 0, -np.inf]]).astype(
+            np.float32
+        )
+        out1_np = np_nan_to_num(x_np)
+        out2_np = np_nan_to_num(x_np, 1.0)
+        out3_np = np_nan_to_num(x_np, 1.0, 9.0)
+        out4_np = np_nan_to_num(x_np, 1.0, 9.0, -12.0)
+        paddle.enable_static()
+        with paddle.static.program_guard(paddle.static.Program()):
+            x = paddle.fluid.data('X', x_np.shape)
+            out1 = paddle.nan_to_num(x)
+            out2 = paddle.nan_to_num(x, 1.0)
+            out3 = paddle.nan_to_num(x, 1.0, 9.0)
+            out4 = paddle.nan_to_num(x, 1.0, 9.0, -12.0)
+            exe = paddle.static.Executor(self.place)
+            res = exe.run(feed={'X': x_np}, fetch_list=[out1, out2, out3, out4])
+
+        self.assertTrue(np.allclose(out1_np, res[0]))
+        self.assertTrue(np.allclose(out2_np, res[1]))
+        self.assertTrue(np.allclose(out3_np, res[2]))
+        self.assertTrue(np.allclose(out4_np, res[3]))
+
+    def test_dygraph(self):
+
+        paddle.disable_static(place=self.place)
+
+        with paddle.fluid.dygraph.guard():
+            # NOTE(tiancaishaonvjituizi): float64 input fails the test
+            x_np = np.array([[1, np.nan, -2], [np.inf, 0, -np.inf]]).astype(
+                np.float32
+                # np.float64
+            )
+            x_tensor = paddle.to_tensor(x_np, stop_gradient=False)
+
+            out_tensor = paddle.nan_to_num(x_tensor)
+            out_np = np_nan_to_num(x_np)
+            self.assertTrue(np.allclose(out_tensor.numpy(), out_np))
+
+            out_tensor = paddle.nan_to_num(x_tensor, 1.0, None, None)
+            out_np = np_nan_to_num(x_np, 1, None, None)
+            self.assertTrue(np.allclose(out_tensor.numpy(), out_np))
+
+            out_tensor = paddle.nan_to_num(x_tensor, 1.0, 2.0, None)
+            out_np = np_nan_to_num(x_np, 1, 2, None)
+            self.assertTrue(np.allclose(out_tensor.numpy(), out_np))
+
+            out_tensor = paddle.nan_to_num(x_tensor, 1.0, None, -10.0)
+            out_np = np_nan_to_num(x_np, 1, None, -10)
+            self.assertTrue(np.allclose(out_tensor.numpy(), out_np))
+
+            out_tensor = paddle.nan_to_num(x_tensor, 1.0, 100.0, -10.0)
+            out_np = np_nan_to_num(x_np, 1, 100, -10)
+            self.assertTrue(np.allclose(out_tensor.numpy(), out_np))
+
+        paddle.enable_static()
+
+    def test_check_grad(self):
+        paddle.disable_static(place=self.place)
+        x_np = np.array([[1, np.nan, -2], [np.inf, 0, -np.inf]]).astype(
+            np.float32
+        )
+        x_tensor = paddle.to_tensor(x_np, stop_gradient=False)
+
+        y = paddle.nan_to_num(x_tensor)
+        dx = paddle.grad(y, x_tensor)[0].numpy()
+
+        np_grad = np_nan_to_num_grad(x_np, np.ones_like(x_np))
+        self.assertTrue(np.allclose(np_grad, dx))
+
+        paddle.enable_static()
+
+
+# class BaseTestCases:
+#
+#     class BaseOpTest(OpTest):
+#
+#         def setUp(self):
+#             self.op_type = "nan_to_num"
+#             input = np.arange(100, dtype=np.float64)
+#             input[5] = np.nan
+#             input[29] = np.inf
+#             input[97] = -np.inf
+#             self.inputs = {'X': input}
+#             self.attrs = self._attrs()
+#             self.outputs = {
+#                 'Out': np_nan_to_num_op(self.inputs['X'], **self.attrs)
+#             }
+#             paddle.enable_static()
+#
+#         def test_check_output(self):
+#             self.check_output()
+#
+#         def test_check_grad(self):
+#             input = self.inputs['X']
+#             dout = np.ones_like(input) / input.size
+#             self.check_grad(
+#                 ['X'],
+#                 'Out',
+#                 user_defined_grads=[np_nan_to_num_grad(self.inputs['X'], dout)])
+#
+#         def _attrs(self):
+#             raise NotImplementedError()
+#
+#
+# class TestNanToNumOp1(BaseTestCases.BaseOpTest):
+#
+#     def _attrs(self):
+#         return {
+#             'nan': 0.0,
+#             'replace_posinf_with_max': True,
+#             'posinf': -1,
+#             'replace_neginf_with_min': True,
+#             'neginf': -10
+#         }
+#
+#
+# class TestNanToNumOp2(BaseTestCases.BaseOpTest):
+#
+#     def _attrs(self):
+#         return {
+#             'nan': 2.0,
+#             'replace_posinf_with_max': False,
+#             'posinf': -1,
+#             'replace_neginf_with_min': True,
+#             'neginf': -10
+#         }
+#
+#
+# class TestNanToNumOp3(BaseTestCases.BaseOpTest):
+#
+#     def _attrs(self):
+#         return {
+#             'nan': 0.0,
+#             'replace_posinf_with_max': False,
+#             'posinf': -1,
+#             'replace_neginf_with_min': False,
+#             'neginf': -10
+#         }
+
+if __name__ == "__main__":
+    unittest.main()

python/paddle/tensor/__init__.py

@@ -169,6 +169,7 @@
 from .math import square  # noqa: F401
 from .math import stanh  # noqa: F401
 from .math import sum  # noqa: F401
+from .math import nan_to_num  # noqa: F401
 from .math import nansum  # noqa: F401
 from .math import nanmean  # noqa: F401
 from .math import count_nonzero  # noqa: F401
@@ -350,6 +351,7 @@
     'square',
     'stanh',
     'sum',
+    'nan_to_num',
     'nansum',
     'nanmean',
     'count_nonzero',

python/paddle/tensor/math.py

@@ -1364,6 +1364,54 @@ def sum(x, axis=None, dtype=None, keepdim=False, name=None):
     return out
 
 
+def nan_to_num(x, nan=0.0, posinf=None, neginf=None, name=None):
+    """
+    Replaces NaN, positive infinity, and negative infinity values in input tensor.
+
+    Args:
+        x (Tensor): An N-D Tensor, the data type is float32, float64.
+        nan (float, optional): the value to replace NaNs with. Default is 0.
+        posinf (float, optional): if a Number, the value to replace positive infinity values with. If None, positive infinity values are replaced with the greatest finite value representable by input’s dtype. Default is None.
+        neginf (float, optional): if a Number, the value to replace negative infinity values with. If None, negative infinity values are replaced with the lowest finite value representable by input’s dtype. Default is None.
+        name (str, optional): Name for the operation (optional, default is None). For more information, please refer to :ref:`api_guide_Name`.
+
+    Returns:
+        Tensor: Results of nan_to_num operation input Tensor ``x``.
+
+    Examples:
+        .. code-block:: python
+
+            import paddle
+
+            x = paddle.to_tensor([float('nan'), 0.3, float('+inf'), float('-inf')], dtype='float32')
+            out1 = paddle.nan_to_num(x)  # [0, 0.3, 3.4028235e+38, -3.4028235e+38]
+            out2 = paddle.nan_to_num(x, nan=1)  # [1, 0.3, 3.4028235e+38, -3.4028235e+38]
+            out3 = paddle.nan_to_num(x, posinf=5)  # [0, 0.3, 5, -3.4028235e+38]
+            out4 = paddle.nan_to_num(x, nan=10, neginf=-99)  # [10, 0.3, 3.4028235e+38, -99]
+    """
+    # NOTE(tiancaishaonvjituizi): it seems that paddle handles the dtype of python float number
+    # incorrectly, so we have to explicitly contruct tensors here
+    posinf_value = paddle.full_like(x, float("+inf"))
+    neginf_value = paddle.full_like(x, float("-inf"))
+    nan = paddle.full_like(x, nan)
+    assert x.dtype in [paddle.float32, paddle.float64]
+    is_float32 = x.dtype == paddle.float32
+    if posinf is None:
+        posinf = (
+            np.finfo(np.float32).max if is_float32 else np.finfo(np.float64).max
+        )
+    posinf = paddle.full_like(x, posinf)
+    if neginf is None:
+        neginf = (
+            np.finfo(np.float32).min if is_float32 else np.finfo(np.float64).min
+        )
+    neginf = paddle.full_like(x, neginf)
+    x = paddle.where(paddle.isnan(x), nan, x)
+    x = paddle.where(x == posinf_value, posinf, x)
+    x = paddle.where(x == neginf_value, neginf, x)
+    return x
+
+
 def nansum(x, axis=None, dtype=None, keepdim=False, name=None):
     """
     Computes the sum of tensor elements over the given axis, treating Not a Numbers (NaNs) as zero.