mindspore.Tensor.scatter_add

Tensor.scatter_add(dim, index, src) → Tensor

Add all elements in src to the index specified by index to self along dimension specified by dim. It takes three inputs self, src and index of the same rank r >= 1.

For a 3-D tensor, the operation updates input as follows:

self[index[i][j][k]][j][k] += src[i][j][k]  # if dim == 0

self[i][index[i][j][k]][k] += src[i][j][k]  # if dim == 1

self[i][j][index[i][j][k]] += src[i][j][k]  # if dim == 2

Note

The rank of this tensor self must be at least 1.

Parameters

dim (int) – Which dim to scatter. Accepted range is [-r, r) where r = rank(self).
index (Tensor) – The index of self to do scatter operation whose data type must be int32 or int64. Same rank as self. Except for the dimension specified by dim, the size of each dimension of index must be less than or equal to the size of the corresponding dimension of self.
src (Tensor) – The tensor doing the scatter operation with self, has the same type as self and the size of each dimension must be greater than or equal to that of index.

Returns

Tensor, has the same shape and type as self.

Raises

TypeError – If index is neither int32 nor int64.
ValueError – If anyone of the rank among self, index and src is less than 1.
ValueError – If the rank of self, index and src is not the same.
ValueError – The size of any dimension of index except the dimension specified by dim is greater than the size of the corresponding dimension of self.
ValueError – If the size of any dimension of src is less than that of index.

Supported Platforms:: Ascend

Examples

>>> import numpy as np
>>> import mindspore as ms
>>> from mindspore import Tensor, ops
>>> input = Tensor(np.array([[1, 2, 3, 4, 5]]), dtype=ms.float32)
>>> src = Tensor(np.array([[8, 8]]), dtype=ms.float32)
>>> index = Tensor(np.array([[2, 4]]), dtype=ms.int64)
>>> out = ops.function.array_func.scatter_add_ext(input=input, dim=1, index=index, src=src)
>>> print(out)
[[1. 2. 11. 4. 13.]]
>>> input = Tensor(np.zeros((5, 5)), dtype=ms.float32)
>>> src = Tensor(np.array([[1, 2, 3], [4, 5, 6], [7, 8, 9]]), dtype=ms.float32)
>>> index = Tensor(np.array([[0, 0, 0], [2, 2, 2], [4, 4, 4]]), dtype=ms.int64)
>>> out = ops.function.array_func.scatter_add_ext(input=input, dim=0, index=index, src=src)
>>> print(out)
[[1. 2. 3. 0. 0.]
 [0. 0. 0. 0. 0.]
 [4. 5. 6. 0. 0.]
 [0. 0. 0. 0. 0.]
 [7. 8. 9. 0. 0.]]
>>> input = Tensor(np.zeros((5, 5)), dtype=ms.float32)
>>> src = Tensor(np.array([[1, 2, 3], [4, 5, 6], [7, 8, 9]]), dtype=ms.float32)
>>> index = Tensor(np.array([[0, 2, 4], [0, 2, 4], [0, 2, 4]]), dtype=ms.int64)
>>> out = ops.function.array_func.scatter_add_ext(input=input, dim=1, index=index, src=src)
>>> print(out)
[[1. 0. 2. 0. 3.]
 [4. 0. 5. 0. 6.]
 [7. 0. 8. 0. 9.]
 [0. 0. 0. 0. 0.]
 [0. 0. 0. 0. 0.]]

Tensor.scatter_add(indices, updates) → Tensor

Creates a new tensor by adding the values from the positions in self 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. This operation is almost equivalent to using ScatterNdAdd, 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 Examples.

\[output\left [indices \right ] = input\_x + update\]

Note

The dimension of this tensor self must be no less than indices.shape[-1].

If some values of the indices are out of bound:

On GPU, if some values of the indices are out of bound, instead of raising an index error, the corresponding updates will not be updated to self tensor.
On CPU, if some values of the indices are out of bound, raising an index error.
On Ascend, out of bound checking is not supported, if some values of the indices are out of bound, unknown errors may be caused.