mindspore.ops.ScatterAdd
- class mindspore.ops.ScatterAdd(use_locking=False)[源代码]
根据指定更新值和输入索引通过加法运算更新输入数据的值。
对于 indices.shape 的每个 i, …, j :
\[\text{input_x}[\text{indices}[i, ..., j], :] \mathrel{+}= \text{updates}[i, ..., j, :]\]输入的 input_x 和 updates 遵循隐式类型转换规则,以确保数据类型一致。如果数据类型不同,则低精度数据类型将转换为高精度的数据类型。当参数的数据类型需要转换时,则会抛出RuntimeError异常。
说明
这是一个运行即更新的算子。因此, input_x 在运算完成后即更新。
- 参数:
use_locking (bool) - 是否启用锁保护。如果为
True
,则 input_x 将受到锁的保护。否则计算结果是未定义的。默认值:False
。
- 输入:
input_x (Union[Parameter, Tensor]) - ScatterAdd的输入,数据类型为Parameter或Tensor。
indices (Tensor) - 指定相加操作的索引,数据类型为mindspore.int32或者mindspore.int64。
updates (Tensor) - 指定与 input_x 相加操作的Tensor,数据类型与 input_x 相同,shape为 indices.shape + x.shape[1:] 。
- 输出:
Tensor,更新后的 input_x ,shape和数据类型与 input_x 相同。
- 异常:
TypeError - use_locking 不是bool。
TypeError - indices 不是int32或者int64。
ValueError - updates 的shape不等于 indices.shape + x.shape[1:] 。
RuntimeError - 当 input_x 和 updates 类型不一致,需要进行类型转换时,如果 updates 不支持转成 input_x 需要的数据类型,就会报错。
- 支持平台:
Ascend
GPU
CPU
样例:
>>> import mindspore >>> import numpy as np >>> from mindspore import Tensor, ops, Parameter >>> input_x = Parameter(Tensor(np.array([[0.0, 0.0, 0.0], [0.0, 0.0, 0.0]]), mindspore.float32), name="x") >>> indices = Tensor(np.array([[0, 1], [1, 1]]), mindspore.int32) >>> updates = Tensor(np.ones([2, 2, 3]), mindspore.float32) >>> scatter_add = ops.ScatterAdd() >>> output = scatter_add(input_x, indices, updates) >>> print(output) [[1. 1. 1.] [3. 3. 3.]] >>> # for input_x will be updated after the operation is completed. input_x need to be re-initialized. >>> input_x = Parameter(Tensor(np.array([[0.0, 0.0, 0.0], [0.0, 0.0, 0.0]]), mindspore.float32), name="x") >>> # for indices = [[0, 1], [1, 1]] >>> # step 1: [0, 1] >>> # input_x[0] = [0.0, 0.0, 0.0] + [1.0, 1.0, 1.0] = [1.0, 1.0, 1.0] >>> # input_x[1] = [0.0, 0.0, 0.0] + [3.0, 3.0, 3.0] = [3.0, 3.0, 3.0] >>> # step 2: [1, 1] >>> # input_x[1] = [3.0, 3.0, 3.0] + [7.0, 7.0, 7.0] = [10.0, 10.0, 10.0] >>> # input_x[1] = [10.0, 10.0, 10.0] + [9.0, 9.0, 9.0] = [19.0, 19.0, 19.0] >>> indices = Tensor(np.array([[0, 1], [1, 1]]), mindspore.int32) >>> updates = Tensor(np.array([[[1.0, 1.0, 1.0], [3.0, 3.0, 3.0]], ... [[7.0, 7.0, 7.0], [9.0, 9.0, 9.0]]]), mindspore.float32) >>> scatter_add = ops.ScatterAdd() >>> output = scatter_add(input_x, indices, updates) >>> print(output) [[ 1. 1. 1.] [19. 19. 19.]] >>> # for input_x will be updated after the operation is completed. input_x need to be re-initialized. >>> input_x = Parameter(Tensor(np.array([[0.0, 0.0, 0.0], [0.0, 0.0, 0.0]]), mindspore.float32), name="x") >>> # for indices = [[1, 0], [1, 1]] >>> # step 1: [1, 0] >>> # input_x[0] = [0.0, 0.0, 0.0] + [3.0, 3.0, 3.0] = [3.0, 3.0, 3.0] >>> # input_x[1] = [0.0, 0.0, 0.0] + [1.0, 1.0, 1.0] = [1.0, 1.0, 1.0] >>> # step 2: [1, 1] >>> # input_x[1] = [1.0, 1.0, 1.0] + [7.0, 7.0, 7.0] = [8.0, 8.0, 8.0] >>> # input_x[1] = [8.0, 8.0, 8.0] + [9.0, 9.0, 9.0] = [17.0, 17.0, 17.0] >>> indices = Tensor(np.array([[1, 0], [1, 1]]), mindspore.int32) >>> updates = Tensor(np.array([[[1.0, 1.0, 1.0], [3.0, 3.0, 3.0]], ... [[7.0, 7.0, 7.0], [9.0, 9.0, 9.0]]]), mindspore.float32) >>> scatter_add = ops.ScatterAdd() >>> output = scatter_add(input_x, indices, updates) >>> print(output) [[ 3. 3. 3.] [17. 17. 17.]] >>> # for input_x will be updated after the operation is completed. input_x need to be re-initialized. >>> input_x = Parameter(Tensor(np.array([[0.0, 0.0, 0.0], [0.0, 0.0, 0.0]]), mindspore.float32), name="x") >>> # for indices = [[0, 1], [0, 1]] >>> # step 1: [0, 1] >>> # input_x[0] = [0.0, 0.0, 0.0] + [1.0, 1.0, 1.0] = [1.0, 1.0, 1.0] >>> # input_x[1] = [0.0, 0.0, 0.0] + [3.0, 3.0, 3.0] = [3.0, 3.0, 3.0] >>> # step 2: [0, 1] >>> # input_x[0] = [1.0, 1.0, 1.0] + [7.0, 7.0, 7.0] = [8.0, 8.0, 8.0] >>> # input_x[1] = [3.0, 3.0, 3.0] + [9.0, 9.0, 9.0] = [12.0, 12.0, 12.0] >>> indices = Tensor(np.array([[0, 1], [0, 1]]), mindspore.int32) >>> updates = Tensor(np.array([[[1.0, 1.0, 1.0], [3.0, 3.0, 3.0]], ... [[7.0, 7.0, 7.0], [9.0, 9.0, 9.0]]]), mindspore.float32) >>> scatter_add = ops.ScatterAdd() >>> output = scatter_add(input_x, indices, updates) >>> print(output) [[ 8. 8. 8.] [12. 12. 12.]]