mindspore.Tensor.scatter_

View Source On Gitee
Tensor.scatter_(dim, index, src) → Tensor

Update the value in src to update self according to the specified index.

Index the dimension self selected by dim using index , traverse the other dimensions in sequence, update the value of src to self , and return self .

This operator is the inverse of the in-place version of mindspore.Tensor.gather() .

This operation provides another three overloads to support parameter reduce and scalar value.

Here's an example using a 3-dimension tensor.

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

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

Warning

  • If multiple indexes point to the same position in self , the final value of this position in self is uncertain.

  • On Ascend, behavior is unpredictable when the value of index is not in the range [-self.shape[dim], self.shape[dim]) in forward.

  • This is an experimental API that is subject to change or deletion.

Note

The inverse gradient from self to src can be calculated only when the shape of src is the same as that of index.

Parameters

  • dim (int) – Which axis to scatter. Accepted range is [-r, r) where r is the rank of self .

  • index (Tensor) – The index to access self on the target axis specified by dim whose dtype must be int32 or int64. If it is an empty Tensor, no operations is performed and directly returns self . Otherwise, its rank must be the same as self and the value range of each element must be [-s, s) where s is the size of self along axis dim .

  • src (Tensor) – The data to doing the update operation with self . It should have the same dtype and rank as self .

Returns

Tensor, the modified self itself.

Raises

  • TypeError – If type of self , index or src is unsupported.

  • RuntimeError – If dim is out of the range [-r, r) .

  • RuntimeError – If rank of self is larger than 8.

  • RuntimeError – If dtype of tensor self , index or src is unsupported.

  • RuntimeError – If dtype of self is not equal to the dtype of src .

  • RuntimeError – If self , index, or src have different ranks and index is not an empty tensor.

  • RuntimeError – If there is a dimension d that makes index.size(d) > src.size(d) .

  • RuntimeError – If there is a dimension d that makes index.size(d) > self.size(d) .

Supported Platforms:

Ascend

Examples

>>> from mindspore import Tensor, int64, float32
>>> this_tensor = Tensor([[1, 2], [3, 4]], dtype=float32)
>>> index = Tensor([[1, 0], [1, 0]], dtype=int64)
>>> src = Tensor([[4, 3], [2, 1]], dtype=float32)
>>> this_tensor.scatter_(1, index, src)
>>> print(this_tensor)
[[3., 4.],
 [1., 2.]]
Tensor.scatter_(dim, index, src, *, reduce) → Tensor

Update the value in src to update self according to the specified index.

Using the operation specified by reduce to index the dimension self selected by dim using index , traverse the other dimensions in sequence, accumulate or multiply the value of src to self , and return self .

This operator is the inverse of the in-place version of mindspore.Tensor.gather() .

Expect that the replacement operation changes to accumulation or multiplication based on the parameter reduce, other operations are the same as the overloaded function that accept src without the parameter reduce .

Here's an example using a 3-dimension tensor.

self[index[i][j][k]][j][k] += src[i][j][k]  # if dim == 0, reduce == "add"

self[i][j][index[i][j][k]] *= src[i][j][k]  # if dim == 2, reduce == "multiply"

Warning

  • If multiple indexes point to the same position in self , the final value of this position in self is uncertain.

  • On Ascend, behavior is unpredictable when the value of index is not in the range [-self.shape[dim], self.shape[dim]) in forward.

  • This is an experimental API that is subject to change or deletion.

Note

This overload function does not support reverse gradient calculation and will return zeros if calculate gradient.

Parameters

  • dim (int) – Which axis to scatter. Accepted range is [-r, r) where r is the rank of self .

  • index (Tensor) – The index to access self on the target axis specified by dim whose dtype must be int32 or int64. If it is an empty Tensor, no operations is performed and directly returns self . Otherwise, its rank must be the same as self and the value range of each element must be [-s, s) where s is the size of self along axis dim .

  • src (Tensor) – The data to doing the accumulate or multiply operation with self . It should have the same dtype and rank as self .

Keyword Arguments

reduce (str) – Reduce operation, supports "add" and "multiply" . When reduce is "add" , src is accumulated to input base on index . When reduce is "multiply" , src is multiplied to input base on index .

Returns

Tensor, the modified self itself.

Raises

  • TypeError – If type of self , index or src is unsupported.

  • ValueError – If reduce is a str but not "add" or "multiply" .

  • RuntimeError – If dim is out of the range [-r, r) .

  • RuntimeError – If rank of self is larger than 8.

  • RuntimeError – If dtype of tensor self , index or src is unsupported.

  • RuntimeError – If dtype of self is not equal to the dtype of src .

  • RuntimeError – If self , index, or src have different ranks and index is not an empty tensor.

  • RuntimeError – If there is a dimension d that makes index.size(d) > src.size(d) .

  • RuntimeError – If there is a dimension d that makes index.size(d) > self.size(d) .

Supported Platforms:

Ascend

Examples

>>> from mindspore import Tensor, int64, float32
>>> this_tensor = Tensor([[1, 2], [3, 4]], dtype=float32)
>>> index = Tensor([[1, 0], [1, 0]], dtype=int64)
>>> src = Tensor([[4, 3], [2, 1]], dtype=float32)
>>> this_tensor.scatter_(1, index, src, reduce='add')
>>> print(this_tensor)
[[4., 6.],
 [4., 6.]]
Tensor.scatter_(dim, index, value) → Tensor

Update the value value to update self according to the specified index.

Index the dimension self selected by dim using index , traverse the other dimensions in sequence, update the value value to self , and return self .

This operator is the inverse of the in-place version of mindspore.Tensor.gather() .

It can be considered that after the value is broadcasted as a Tensor whose shape and dtype are consistent with self , other operations are the same as the overloaded function that accept src without the parameter reduce .

Here's an example using a 3-dimension tensor.

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

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

Warning

  • If multiple indexes point to the same position in self , the final value of this position in self is uncertain.

  • On Ascend, behavior is unpredictable when the value of index is not in the range [-self.shape[dim], self.shape[dim]) in forward.

  • This is an experimental API that is subject to change or deletion.

Parameters

  • dim (int) – Which axis to scatter. Accepted range is [-r, r) where r is the rank of self .

  • index (Tensor) – The index to access self on the target axis specified by dim whose dtype must be int32 or int64. If it is an empty Tensor, no operations is performed and directly returns self . Otherwise, its rank must be the same as self and the value range of each element must be [-s, s) where s is the size of self along axis dim .

  • value (int, float, bool) – The data to doing the update operation with self . It can be considered as being broadcasted into a Tensor whose shape and dtype are the same as self , and then be regarded as src for calculation.

Returns

Tensor, the modified self itself.

Raises

  • TypeError – If type of self , index or value is unsupported.

  • RuntimeError – If dim is out of the range [-r, r) .

  • RuntimeError – If rank of self is larger than 8.

  • RuntimeError – If dtype of tensor self or index is unsupported.

  • RuntimeError – If index is not an empty tensor and its rank is different from the rank of self .

  • RuntimeError – If there is a dimension d that makes index.size(d) > self.size(d) .

Supported Platforms:

Ascend

Examples

>>> from mindspore import Tensor, int64, float32
>>> this_tensor = Tensor([[1, 2], [3, 4]], dtype=float32)
>>> index = Tensor([[0], [1]], dtype=int64)
>>> this_tensor.scatter_(0, index, 10)
>>> print(this_tensor)
[[10., 2.],
 [10., 4.]]
Tensor.scatter_(dim, index, value, *, reduce) → Tensor

Update the value value to update self according to the specified index.

Using the operation specified by reduce to index the dimension self selected by dim using index , traverse the other dimensions in sequence, accumulate or multiply the value value to self , and return self .

This operator is the inverse of the in-place version of mindspore.Tensor.gather() .

Expect that the replacement operation changes to accumulation or multiplication based on the parameter reduce, other operations are the same as the overloaded function that accept value without the parameter reduce .

Here's an example using a 3-dimension tensor.

self[i][index[i][j][k]][k] += value  # if dim == 1, reduce == "add"

self[i][j][index[i][j][k]] *= value  # if dim == 2, reduce == "multiply"

Warning

  • If multiple indexes point to the same position in self , the final value of this position in self is uncertain.

  • On Ascend, behavior is unpredictable when the value of index is not in the range [-self.shape[dim], self.shape[dim]) in forward.

  • This is an experimental API that is subject to change or deletion.

Note

This overload function does not support reverse gradient calculation and will return zeros if calculate gradient.

Parameters

  • dim (int) – Which axis to scatter. Accepted range is [-r, r) where r is the rank of self .

  • index (Tensor) – The index to access self on the target axis specified by dim whose dtype must be int32 or int64. If it is an empty Tensor, no operations is performed and directly returns self . Otherwise, its rank must be the same as self and the value range of each element must be [-s, s) where s is the size of self along axis dim .

  • value (int, float, bool) – The data to doing the accumulate or multiply operation with self . It can be considered as being broadcasted into a Tensor whose shape and dtype are the same as self , and then be regarded as src for calculation.

Keyword Arguments

reduce (str) – Reduce operation, supports "add" and "multiply" . When reduce is "add" , value is accumulated to input base on index . When reduce is "multiply" , value is multiplied to input base on index .

Returns

Tensor, the modified self itself.

Raises

  • TypeError – If type of self , index or value is unsupported.

  • ValueError – If reduce is a str but not "add" or "multiply" .

  • RuntimeError – If dim is out of the range [-r, r) .

  • RuntimeError – If rank of self is larger than 8.

  • RuntimeError – If dtype of tensor self or index is unsupported.

  • RuntimeError – If index is not an empty tensor and its rank is different from the rank of self .

  • RuntimeError – If there is a dimension d that makes index.size(d) > self.size(d) .

Supported Platforms:

Ascend

Examples

>>> from mindspore import Tensor, int64, float32
>>> this_tensor = Tensor([[1, 2], [3, 4]], dtype=float32)
>>> index = Tensor([[0], [1]], dtype=int64)
>>> this_tensor.scatter_(0, index, 3, reduce="multiply")
>>> print(this_tensor)
[[3., 2.],
 [9., 4.]]