Add custom op interface to replace expm1, log1p and log

mindspore/nn/probability/bijector/power_transform.py

@@ -17,6 +17,7 @@ from mindspore.ops import operations as P
 from mindspore._checkparam import Validator as validator
 from mindspore._checkparam import Rel
 from ..distribution._utils.utils import CheckTensor
+from ..distribution._utils.custom_ops import log_by_step, log1p_by_step, expm1_by_step
 from .bijector import Bijector
 
 class PowerTransform(Bijector):
@@ -59,24 +60,12 @@ class PowerTransform(Bijector):
         self._power = power
         self.pow = P.Pow()
         self.exp = P.Exp()
-        self.log = P.Log()
-        self.log1p = self._log1p_by_step
-        self.expm1 = self._expm1_by_step
+        self.log = log_by_step
+        self.log1p = log1p_by_step
+        self.expm1 = expm1_by_step
 
         self.checktensor = CheckTensor()
 
-    def _log1p_by_step(self, x):
-        """
-        Log1p ops on GPU device or when device_target == GPU.
-        """
-        return self.log(x + 1.0)
-
-    def _expm1_by_step(self, x):
-        """
-        Expm1 ops on GPU device or when device_target == GPU.
-        """
-        return self.exp(x) - 1.0
-
     @property
     def power(self):
         return self._power

mindspore/nn/probability/bijector/scalar_affine.py

@@ -16,6 +16,7 @@
 from mindspore.ops import operations as P
 from mindspore._checkparam import Validator as validator
 from ..distribution._utils.utils import cast_to_tensor, CheckTensor
+from ..distribution._utils.custom_ops import log_by_step
 from .bijector import Bijector
 
 class ScalarAffine(Bijector):
@@ -66,7 +67,7 @@ class ScalarAffine(Bijector):
             param=param)
 
         self.abs = P.Abs()
-        self.log = P.Log()
+        self.log = log_by_step
 
         self.checktensor = CheckTensor()
 

mindspore/nn/probability/bijector/softplus.py

@@ -19,6 +19,7 @@ from mindspore.common import dtype as mstype
 from mindspore.nn.layer.activation import LogSigmoid
 from mindspore._checkparam import Validator as validator
 from ..distribution._utils.utils import cast_to_tensor, CheckTensor
+from ..distribution._utils.custom_ops import log_by_step, expm1_by_step
 from .bijector import Bijector
 
 class Softplus(Bijector):
@@ -60,12 +61,12 @@ class Softplus(Bijector):
 
         self.abs = P.Abs()
         self.exp = P.Exp()
-        self.expm1 = self._expm1_by_step
+        self.log = log_by_step
+        self.expm1 = expm1_by_step
         self.fill = P.Fill()
         self.greater = P.Greater()
         self.less = P.Less()
         self.log_sigmoid = LogSigmoid()
-        self.log = P.Log()
         self.logicalor = P.LogicalOr()
         self.select = P.Select()
         self.shape = P.Shape()
@@ -76,12 +77,6 @@ class Softplus(Bijector):
         self.checktensor = CheckTensor()
         self.threshold = np.log(np.finfo(np.float32).eps) + 1
 
-    def _expm1_by_step(self, x):
-        """
-        Expm1 ops under GPU context.
-        """
-        return self.exp(x) - 1.0
-
     def _softplus(self, x):
         too_small = self.less(x, self.threshold)
         too_large = self.greater(x, -self.threshold)

mindspore/nn/probability/distribution/_utils/__init__.py

@@ -16,6 +16,7 @@
 Distribution operation utility functions.
 """
 from .utils import *
+from .custom_ops import *
 
 __all__ = [
     'convert_to_batch',
@@ -27,4 +28,7 @@ __all__ = [
     'check_scalar_from_param',
     'check_prob',
     'check_type',
+    'log_by_step',
+    'log1p_by_step',
+    'expm1_by_step',
 ]

mindspore/nn/probability/distribution/_utils/custom_ops.py

@@ -0,0 +1,48 @@
+# Copyright 2020 Huawei Technologies Co., Ltd
+#
+# 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.
+# ============================================================================
+"""Utitly functions to help distribution class."""
+import numpy as np
+from mindspore.ops import operations as P
+
+def log_by_step(input_x):
+    """
+    Log op on Ascend is calculated as log(abs(x)).
+    Fix this with putting negative values as nan.
+    """
+    select = P.Select()
+    log = P.Log()
+    lessequal = P.LessEqual()
+    fill = P.Fill()
+    dtype = P.DType()
+    shape = P.Shape()
+
+    nonpos_x = lessequal(input_x, 0.0)
+    log_x = log(input_x)
+    nan = fill(dtype(input_x), shape(input_x), np.nan)
+    result = select(nonpos_x, nan, log_x)
+    return result
+
+def log1p_by_step(x):
+    """
+    Log1p ops on GPU device or when device_target == GPU.
+    """
+    return log_by_step(x + 1.0)
+
+def expm1_by_step(input_x):
+    """
+    Expm1 ops under GPU context.
+    """
+    exp = P.Exp()
+    return exp(input_x) - 1.0

mindspore/nn/probability/distribution/bernoulli.py

@@ -19,6 +19,7 @@ from mindspore.ops import composite as C
 from .distribution import Distribution
 from ._utils.utils import cast_to_tensor, check_prob, check_type, check_distribution_name, raise_none_error
 from ._utils.utils import CheckTensor, CheckTuple
+from ._utils.custom_ops import log_by_step
 
 class Bernoulli(Distribution):
     """
@@ -116,7 +117,7 @@ class Bernoulli(Distribution):
         self.exp = P.Exp()
         self.floor = P.Floor()
         self.fill = P.Fill()
-        self.log = P.Log()
+        self.log = log_by_step
         self.less = P.Less()
         self.shape = P.Shape()
         self.select = P.Select()

mindspore/nn/probability/distribution/exponential.py

@@ -21,6 +21,7 @@ from .distribution import Distribution
 from ._utils.utils import cast_to_tensor, check_greater_zero, check_type, check_distribution_name,\
                           raise_none_error
 from ._utils.utils import CheckTensor, CheckTuple
+from ._utils.custom_ops import log_by_step
 
 class Exponential(Distribution):
     """
@@ -119,7 +120,7 @@ class Exponential(Distribution):
         self.exp = P.Exp()
         self.fill = P.Fill()
         self.less = P.Less()
-        self.log = P.Log()
+        self.log = log_by_step
         self.select = P.Select()
         self.shape = P.Shape()
         self.sqrt = P.Sqrt()

mindspore/nn/probability/distribution/geometric.py

@@ -21,6 +21,7 @@ from .distribution import Distribution
 from ._utils.utils import cast_to_tensor, check_prob, check_type, check_distribution_name,\
                           raise_none_error
 from ._utils.utils import CheckTensor, CheckTuple
+from ._utils.custom_ops import log_by_step
 
 class Geometric(Distribution):
     """
@@ -122,7 +123,7 @@ class Geometric(Distribution):
         self.floor = P.Floor()
         self.issubclass = P.IsSubClass()
         self.less = P.Less()
-        self.log = P.Log()
+        self.log = log_by_step
         self.pow = P.Pow()
         self.select = P.Select()
         self.shape = P.Shape()

mindspore/nn/probability/distribution/normal.py

@@ -21,6 +21,7 @@ from .distribution import Distribution
 from ._utils.utils import convert_to_batch, check_greater_zero, check_type, check_distribution_name,\
                           raise_none_error
 from ._utils.utils import CheckTensor, CheckTuple
+from ._utils.custom_ops import log_by_step, expm1_by_step
 
 class Normal(Distribution):
     """
@@ -119,9 +120,9 @@ class Normal(Distribution):
         self.const = P.ScalarToArray()
         self.erf = P.Erf()
         self.exp = P.Exp()
-        self.expm1 = self._expm1_by_step
+        self.expm1 = expm1_by_step
         self.fill = P.Fill()
-        self.log = P.Log()
+        self.log = log_by_step
         self.shape = P.Shape()
         self.sq = P.Square()
         self.sqrt = P.Sqrt()
@@ -137,12 +138,6 @@ class Normal(Distribution):
             str_info = f'batch_shape = {self._broadcast_shape}'
         return str_info
 
-    def _expm1_by_step(self, x):
-        """
-        Expm1 ops under GPU context.
-        """
-        return self.exp(x) - 1.0
-
     def _check_param(self, mean, sd):
         """
         Check availablity of distribution specific args mean and sd.

mindspore/nn/probability/distribution/transformed_distribution.py

@@ -19,6 +19,7 @@ from mindspore.common import dtype as mstype
 import mindspore.nn as nn
 from .distribution import Distribution
 from ._utils.utils import check_type, raise_not_impl_error
+from ._utils.custom_ops import log_by_step
 
 class TransformedDistribution(Distribution):
     """
@@ -56,7 +57,7 @@ class TransformedDistribution(Distribution):
         self._distribution = distribution
         self._is_linear_transformation = bijector.is_constant_jacobian
         self.exp = P.Exp()
-        self.log = P.Log()
+        self.log = log_by_step
 
     @property
     def bijector(self):

mindspore/nn/probability/distribution/uniform.py

@@ -20,6 +20,7 @@ from .distribution import Distribution
 from ._utils.utils import convert_to_batch, check_greater, check_type, check_distribution_name,\
                           raise_none_error
 from ._utils.utils import CheckTensor, CheckTuple
+from ._utils.custom_ops import log_by_step
 
 class Uniform(Distribution):
     """
@@ -121,7 +122,7 @@ class Uniform(Distribution):
         self.fill = P.Fill()
         self.less = P.Less()
         self.lessequal = P.LessEqual()
-        self.log = P.Log()
+        self.log = log_by_step
         self.logicaland = P.LogicalAnd()
         self.select = P.Select()
         self.shape = P.Shape()