Differences with torch.scatter

The following mapping relationships can be found in this file.

PyTorch APIs	MindSpore APIs
torch.scatter	mindspore.ops.scatter
torch.Tensor.scatter	mindspore.Tensor.scatter

torch.scatter

torch.scatter(input, dim, index, src)

For more information, see torch.scatter.

mindspore.ops.scatter

mindspore.ops.scatter(input, axis, index, src)

For more information, see mindspore.ops.scatter.

Differences

API function of MindSpore is not consistent with that of PyTorch.

PyTorch: For all dimensions d , index.size(d) <= src.size(d) is required, i.e. index can select some or all of the data of src to be scattered into input .

MindSpore: The shape of index must be the same as the shape of src , i.e. all data of src will be scattered into input by index .

Categories	Subcategories	PyTorch	MindSpore	Differences
Parameters	Parameter 1	input	input	Consistent
	Parameter 2	dim	axis	Different parameter names
	Parameter 3	index	index	For MindSpore, the shape of `index` must be the same as the shape of `src` . For PyTorch, `index.size(d) <= src.size(d)` is required for all dimensions `d`
	Parameter 4	src	src	Consistent

Code Example 1

Scatter operation on part of src data.

# PyTorch
import torch
import numpy as np
input = torch.tensor(np.zeros((5, 5)), dtype=torch.float32)
src = torch.tensor(np.array([[1, 2, 3], [4, 5, 6], [7, 8, 9]]), dtype=torch.float32)
index = torch.tensor(np.array([[0, 1], [0, 1], [0, 1]]), dtype=torch.int64)
out = torch.scatter(input=input, dim=1, index=index, src=src)
print(out)
# tensor([[1., 2., 0., 0., 0.],
#         [4., 5., 0., 0., 0.],
#         [7., 8., 0., 0., 0.],
#         [0., 0., 0., 0., 0.],
#         [0., 0., 0., 0., 0.]])

# MindSpore doesn't support this feature currently.

Code Example 2

Scatter operation on all of src data. The output values of singular value decomposition are not unique.

import torch
import numpy as np
input = torch.tensor(np.zeros((5, 5)), dtype=torch.float32)
src = torch.tensor(np.array([[1, 2, 3], [4, 5, 6], [7, 8, 9]]), dtype=torch.float32)
index = torch.tensor(np.array([[0, 1, 2], [0, 1, 2], [0, 1, 2]]), dtype=torch.int64)
out = torch.scatter(input=input, dim=1, index=index, src=src)
print(out)
# tensor([[1., 2., 3., 0., 0.],
#         [4., 5., 6., 0., 0.],
#         [7., 8., 9., 0., 0.],
#         [0., 0., 0., 0., 0.],
#         [0., 0., 0., 0., 0.]])

# MindSpore
import mindspore as ms
import numpy as np
input = ms.Tensor(np.zeros((5, 5)), dtype=ms.float32)
src = ms.Tensor(np.array([[1, 2, 3], [4, 5, 6], [7, 8, 9]]), dtype=ms.float32)
index = ms.Tensor(np.array([[0, 1, 2], [0, 1, 2], [0, 1, 2]]), dtype=ms.int64)
out = ms.ops.scatter(input=input, axis=1, index=index, src=src)
print(out)
# [[1. 2. 3. 0. 0.]
#  [4. 5. 6. 0. 0.]
#  [7. 8. 9. 0. 0.]
#  [0. 0. 0. 0. 0.]
#  [0. 0. 0. 0. 0.]]