!34078 Add tensor&functional interface for TensorScatterSub ops.

Merge pull request !34078 from liangzelang/dev_tensorscatterop_cpu

docs/api/api_python/mindspore.ops.functional.rst

@@ -268,6 +268,7 @@ Array操作
     mindspore.ops.shape
     mindspore.ops.size
     mindspore.ops.tensor_scatter_add
+    mindspore.ops.tensor_scatter_sub
     mindspore.ops.tensor_scatter_div
     mindspore.ops.space_to_batch_nd
     mindspore.ops.tile

docs/api/api_python/mindspore/mindspore.Tensor.rst

@@ -890,6 +890,29 @@ mindspore.Tensor
         - **TypeError** - `indices` 的数据类型既不是int32，也不是int64。
         - **ValueError** - Tensor的shape长度小于 `indices` 的shape的最后一个维度。
 
+    .. py:method:: tensor_scatter_sub(indices, updates)
+
+        根据指定的更新值和输入索引，通过减法进行运算，将结果赋值到输出Tensor中。当同一索引有不同值时，更新的结果将是所有值的总和。此操作几乎等同于使用 :class:`mindspore.ops.ScatterNdSub` ，只是更新后的结果是通过算子output返回，而不是直接原地更新input。
+
+        `indices` 的最后一个轴是每个索引向量的深度。对于每个索引向量， `updates` 中必须有相应的值。`updates` 的shape应该等于 `input_x[indices]` 的shape。有关更多详细信息，请参见使用用例。
+
+        .. note::
+            如果 `indices` 的某些值超出范围，则相应的 `updates` 不会更新到 `input_x` ，而不是抛出索引错误。
+
+        **参数：**
+
+        - **indices** (Tensor) - Tensor的索引，数据类型为int32或int64的。其rank必须至少为2。
+        - **updates** (Tensor) - 指定与本Tensor相减操作的Tensor，其数据类型与该Tensor相同。updates.shape应等于indices.shape[:-1] + self.shape[indices.shape[-1]:]。
+
+        **返回：**
+
+        Tensor，shape和数据类型与原Tensor相同。
+
+        **异常：**
+
+        - **TypeError** - `indices` 的数据类型既不是int32，也不是int64。
+        - **ValueError** - Tensor的shape长度小于 `indices` 的shape的最后一个维度。
+
     .. py:method:: tensor_scatter_div(indices, updates)
 
         根据指定的索引, 通过除法进行计算, 将输出赋值到输出Tensor中。

docs/api/api_python/ops/mindspore.ops.TensorScatterSub.rst

@@ -5,23 +5,5 @@
 
     根据指定的更新值和输入索引，通过减法运算更新输入Tensor的值。当同一个索引有多个不同值时，更新的结果将分别减去这些值。此操作几乎等同于使用 :class:`mindspore.ops.ScatterNdSub` ，只是更新后的结果是通过算子output返回，而不是直接原地更新input。
 
-    `indices` 的最后一个轴是每个索引向量的深度。对于每个索引向量， `updates` 中必须有相应的值。`updates` 的shape应该等于 `input_x[indices]` 的shape。有关更多详细信息，请参见使用用例。
+    更多参考详见 ：func:`mindspore.ops.tensor_scatter_sub`。
-
-    .. note::
-        如果 `indices` 的某些值超出范围，则相应的 `updates` 不会更新为 `input_x` ，而不是抛出索引错误。
-
-    **输入：**
-
-    - **input_x** (Tensor) - 输入Tensor。 `input_x` 的维度必须不小于indices.shape[-1]。
-    - **indices** (Tensor) - 输入Tensor的索引，数据类型为int32或int64的，其rank必须至少为2。
-    - **updates** (Tensor) - 指定与 `input_x` 相减操作的Tensor，其数据类型与输入相同。updates.shape应等于indices.shape[:-1] + input_x.shape[indices.shape[-1]:]。
-
-    **输出：**
-
-    Tensor，shape和数据类型与输入 `input_x` 相同。
-
-    **异常：**
-
-    - **TypeError** - `indices` 的数据类型既不是int32，也不是int64。
-    - **ValueError** - `input_x` 的shape长度小于 `indices` 的shape的最后一个维度。
     

docs/api/api_python/ops/mindspore.ops.func_tensor_scatter_sub.rst

@@ -0,0 +1,26 @@
+mindspore.ops.tensor_scatter_sub
+================================
+
+.. py:function:: mindspore.ops.tensor_scatter_sub(input_x, indices, updates)
+
+    根据指定的更新值和输入索引，通过减法进行运算，将结果赋值到输出Tensor中。当同一索引有不同值时，更新的结果将是所有值的总和。此操作几乎等同于使用 :class:`mindspore.ops.ScatterNdSub` ，只是更新后的结果是通过算子output返回，而不是直接原地更新input。
+
+    `indices` 的最后一个轴是每个索引向量的深度。对于每个索引向量， `updates` 中必须有相应的值。 `updates` 的shape应该等于 `input_x[indices]` 的shape。有关更多详细信息，请参见使用用例。
+
+    .. note::
+        如果 `indices` 的某些值超出范围，则相应的 `updates` 不会更新到 `input_x` ，而不是抛出索引错误。
+
+    **参数：**
+
+    - **input_x** (Tensor) - 输入Tensor。 `input_x` 的维度必须不小于indices.shape[-1]。
+    - **indices** (Tensor) - 输入Tensor的索引，数据类型为int32或int64的。其rank必须至少为2。
+    - **updates** (Tensor) - 指定与 `input_x` 相减操作的Tensor，其数据类型与输入相同。updates.shape应等于indices.shape[:-1] + input_x.shape[indices.shape[-1]:]。
+
+    **返回：**
+
+    Tensor，shape和数据类型与输入 `input_x` 相同。
+
+    **异常：**
+
+    - **TypeError** - `indices` 的数据类型既不是int32，也不是int64。
+    - **ValueError** - `input_x` 的shape长度小于 `indices` 的shape的最后一个维度。

docs/api/api_python_en/mindspore.ops.functional.rst

@@ -267,6 +267,7 @@ Array Operation
     mindspore.ops.shape
     mindspore.ops.size
     mindspore.ops.tensor_scatter_add
+    mindspore.ops.tensor_scatter_sub
     mindspore.ops.tensor_scatter_div
     mindspore.ops.space_to_batch_nd
     mindspore.ops.tile

mindspore/ccsrc/pipeline/jit/resource.cc

@@ -219,6 +219,7 @@ BuiltInTypeMap &GetMethodMap() {
        {"scatter_nd_add", std::string("scatter_nd_add")},                  // scatter_nd_add()
        {"scatter_nd_sub", std::string("scatter_nd_sub")},                  // scatter_nd_sub()
        {"tensor_scatter_add", std::string("tensor_scatter_add")},          // P.TensorScatterAdd()
+       {"tensor_scatter_sub", std::string("tensor_scatter_sub")},          // P.TensorScatterSub()
        {"tensor_scatter_div", std::string("tensor_scatter_div")},          // P.TensorScatterDiv()
        {"lp_norm", std::string("lp_norm")},                                // lp_norm()
        {"trace", std::string("trace")},                                    // P.Eye()

mindspore/python/mindspore/_extends/parse/standard_method.py

@@ -1685,6 +1685,15 @@ def tensor_scatter_add(x, indices, updates):
     return F.tensor_scatter_add(x, indices, updates)
 
 
+def tensor_scatter_sub(x, indices, updates):
+    """
+    Creates a new tensor by subtracting the values from the positions in `x` indicated by
+    `indices`, with values from `updates`. When multiple values are given for the same
+    index, the updated result will be the sum of all values.
+    """
+    return F.tensor_scatter_sub(x, indices, updates)
+
+
 def tensor_sactter_div(input_x, indices, updates):
     """
     Create a new tensor by division the values from the positions in `input_x` indicated by

mindspore/python/mindspore/common/tensor.py

@@ -1989,6 +1989,52 @@ class Tensor(Tensor_):
         self._init_check()
         return tensor_operator_registry.get("tensor_scatter_add")()(self, indices, updates)
 
+    def tensor_scatter_sub(self, indices, updates):
+        """
+        Creates a new tensor by subtracting the values from the positions in self tensor indicated by
+        `indices`, with values from `updates`. When multiple values are provided for the same
+        index, the result of the update will be to subtract these values respectively. This operation is almost
+        equivalent to using :class:`mindspore.ops.ScatterNdSub` , except that the updates are applied on output `Tensor`
+        instead of input `Parameter`.
+
+        The last axis of `indices` is the depth of each index vectors. For each index vector,
+        there must be a corresponding value in `updates`. The shape of `updates` should be
+        equal to the shape of `self[indices]`. For more details, see use cases.
+
+        Note:
+            If some values of the `indices` are out of bound, instead of raising an index error,
+            the corresponding `updates` will not be updated to `input_x`.
+
+        Args:
+            indices (Tensor): The index of input tensor whose data type is int32 or int64.
+                The rank must be at least 2.
+            updates (Tensor): The tensor to update the input tensor, has the same type as input,
+                and updates.shape should be equal to indices.shape[:-1] + self.shape[indices.shape[-1]:].
+
+        Returns:
+            Tensor, has the same shape and type as self tensor.
+
+        Raises:
+            TypeError: If dtype of `indices` is neither int32 nor int64.
+            ValueError: If length of shape of self tensor is less than the last dimension of shape of `indices`.
+
+        Supported Platforms:
+            ``Ascend`` ``GPU`` ``CPU``
+
+        Examples:
+            >>> import numpy as np
+            >>> from mindspore import Tensor
+            >>> x = Tensor(np.array([[-0.1, 0.3, 3.6], [0.4, 0.5, -3.2]]).astype('float32'))
+            >>> indices = Tensor(np.array([[0, 0], [0, 0]]).astype('int32'))
+            >>> updates = Tensor(np.array([1.0, 2.2]).astype('float32'))
+            >>> output = x.tensor_scatter_sub(indices, updates)
+            >>> print(output)
+            [[-3.3000002  0.3        3.6      ]
+            [ 0.4        0.5       -3.2      ]]
+        """
+        self._init_check()
+        return tensor_operator_registry.get('tensor_scatter_sub')()(self, indices, updates)
+
     def fill(self, value):
         """
         Fill the tensor with a scalar value.

mindspore/python/mindspore/ops/function/__init__.py

@@ -23,9 +23,9 @@ from . import array_func, parameter_func, math_func, nn_func
 from .array_func import (unique, eye, matrix_band_part, fill, fill_, tile, size, ones, ones_like, shape, shape_, ger,
                          dyn_shape, rank, reshape, reshape_, tensor_slice, slice, scalar_to_array, scalar_to_tensor,
                          tuple_to_array, expand_dims, transpose, scatter_nd, scatter_nd_add, scatter_nd_sub, gather,
-                         gather_d, gather_nd, scalar_cast,
+                         gather_d, gather_nd, scalar_cast, masked_fill, tensor_scatter_add, tensor_scatter_sub,
-                         masked_fill, tensor_scatter_add, tensor_scatter_div, scatter_max, scatter_min, nonzero,
+                         tensor_scatter_div, scatter_max, scatter_min, nonzero, space_to_batch_nd, range, select)
-                         space_to_batch_nd, range, select)
+
 from .parameter_func import assign, assign_add, assign_sub, index_add
 from .math_func import (addn, absolute, abs, tensor_add, add, neg_tensor, neg, tensor_lt, less, tensor_le, le, lerp,
                         lp_norm, round, tensor_gt, gt, tensor_ge, ge, tensor_sub, sub, tensor_mul, mul, tensor_div, div,

mindspore/python/mindspore/ops/function/array_func.py

@@ -1398,6 +1398,57 @@ def tensor_scatter_add(input_x, indices, updates):
     return tensor_scatter_add_(input_x, indices, updates)
 
 
+tensor_scatter_sub_ = P.TensorScatterSub()
+def tensor_scatter_sub(input_x, indices, updates):
+    """
+    Creates a new tensor by subtracting the values from the positions in `input_x` indicated by
+    `indices`, with values from `updates`. When multiple values are provided for the same
+    index, the result of the update will be to subtract these values respectively. This operation is almost
+    equivalent to using :class:`mindspore.ops.ScatterNdSub` , except that the updates are applied on output `Tensor`
+    instead of input `Parameter`.
+
+    The last axis of `indices` is the depth of each index vectors. For each index vector,
+    there must be a corresponding value in `updates`. The shape of `updates` should be
+    equal to the shape of `input_x[indices]`. For more details, see use cases.
+
+    Note:
+        If some values of the `indices` are out of bound, instead of raising an index error,
+        the corresponding `updates` will not be updated to `input_x`.
+
+    Args:
+        input_x (Tensor): The target tensor. The dimension of input_x must be no less than indices.shape[-1].
+        indices (Tensor): The index of input tensor whose data type is int32 or int64.
+            The rank must be at least 2.
+        updates (Tensor): The tensor to update the input tensor, has the same type as input,
+            and updates.shape should be equal to indices.shape[:-1] + input_x.shape[indices.shape[-1]:].
+
+    Returns:
+        Tensor, has the same shape and type as `input_x`.
+
+    Raises:
+        TypeError: If dtype of `indices` is neither int32 nor int64.
+        ValueError: If length of shape of `input_x` is less than the last dimension of shape of `indices`.
+
+    Supported Platforms:
+        ``Ascend`` ``GPU`` ``CPU``
+
+    Examples:
+        >>> import mindspore
+        >>> import numpy as np
+        >>> from mindspore import Tensor
+        >>> from mindspore import ops
+        >>> input_x = Tensor(np.array([[-0.1, 0.3, 3.6], [0.4, 0.5, -3.2]]), mindspore.float32)
+        >>> indices = Tensor(np.array([[0, 0], [0, 0]]), mindspore.int32)
+        >>> updates = Tensor(np.array([1.0, 2.2]), mindspore.float32)
+        >>> output = ops.tensor_scatter_sub(input_x, indices, updates)
+        >>> print(output)
+        [[-3.3000002  0.3        3.6      ]
+         [ 0.4        0.5       -3.2      ]]
+    """
+
+    return tensor_scatter_sub_(input_x, indices, updates)
+
+
 def space_to_batch_nd(input_x, block_size, paddings):
     r"""
     Divides a tensor's spatial dimensions into blocks and combines the block sizes with the original batch.
@@ -1724,6 +1775,7 @@ __all__ = [
     'scatter_nd_add',
     'scatter_nd_sub',
     'tensor_scatter_add',
+    'tensor_scatter_sub',
     'gather',
     'gather_d',
     'gather_nd',

mindspore/python/mindspore/ops/functional.py

@@ -70,7 +70,6 @@ scatter_update = P.ScatterUpdate()
 tensor_scatter_update = P.TensorScatterUpdate()
 tensor_scatter_min = P.TensorScatterMin()
 tensor_scatter_max = P.TensorScatterMax()
-tensor_scatter_sub = P.TensorScatterSub()
 tensor_scatter_mul = P.TensorScatterMul()
 scatter_nd_update = P.ScatterNdUpdate()
 stack = P.Stack()
@@ -942,6 +941,7 @@ tensor_operator_registry.register('scatter_nd_add', scatter_nd_add)
 tensor_operator_registry.register('scatter_nd_sub', scatter_nd_sub)
 tensor_operator_registry.register('tensor_scatter_update', tensor_scatter_update)
 tensor_operator_registry.register('tensor_scatter_div', tensor_scatter_div)
+tensor_operator_registry.register('tensor_scatter_sub', P.TensorScatterSub)
 tensor_operator_registry.register('tensor_scatter_add', P.TensorScatterAdd)
 tensor_operator_registry.register('lp_norm', lp_norm)
 __all__ = [name for name in dir() if name[0] != "_"]

mindspore/python/mindspore/ops/operations/array_ops.py

@@ -6801,31 +6801,10 @@ class TensorScatterSub(_TensorScatterOp):
     index, the result of the update will be to subtract these values respectively. This operation is almost
     equivalent to using :class:`mindspore.ops.ScatterNdSub` , except that the updates are applied on output `Tensor`
     instead of input `Parameter`.
-
+    Refer to :func:`mindspore.ops.tensor_scatter_sub` for more detail.
-    The last axis of `indices` is the depth of each index vectors. For each index vector,
-    there must be a corresponding value in `updates`. The shape of `updates` should be
-    equal to the shape of `input_x[indices]`. For more details, see use cases.
-
-    Note:
-        If some values of the `indices` are out of bound, instead of raising an index error,
-        the corresponding `updates` will not be updated to `input_x`.
-
-    Inputs:
-        - **input_x** (Tensor) - The target tensor. The dimension of input_x must be no less than indices.shape[-1].
-        - **indices** (Tensor) - The index of input tensor whose data type is int32 or int64.
-          The rank must be at least 2.
-        - **updates** (Tensor) - The tensor to update the input tensor, has the same type as input,
-          and updates.shape should be equal to indices.shape[:-1] + input_x.shape[indices.shape[-1]:].
-
-    Outputs:
-        Tensor, has the same shape and type as `input_x`.
-
-    Raises:
-        TypeError: If dtype of `indices` is neither int32 nor int64.
-        ValueError: If length of shape of `input_x` is less than the last dimension of shape of `indices`.
 
     Supported Platforms:
-        ``GPU``
+        ``Ascend`` ``GPU`` ``CPU``
 
     Examples:
         >>> input_x = Tensor(np.array([[-0.1, 0.3, 3.6], [0.4, 0.5, -3.2]]), mindspore.float32)