Differences with torch.svd

The following mapping relationships can be found in this file.

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

torch.svd

torch.svd(input, some=True, compute_uv=True, *, out=None)

For more information, see torch.svd.

mindspore.ops.svd

mindspore.ops.svd(input, full_matrices=False, compute_uv=True)

For more information, see mindspore.ops.svd.

Differences

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

PyTorch:

If some is True, the method returns the reduced singular value decomposition.
There are always three output values, and the order of the output values is u, s, v.
If compute_uv is False, the returned u and v will be zero-filled matrices.

MindSpore:

If full_matrices is False, the method returns the reduced singular value decomposition.
If compute_uv is False, there is only one output value s.
If compute_uv is True, there are three output values in the order s, u, v.

torch.svd() has been deprecated in PyTorch 1.8.0 and later, and alternative api torch.linalg.svd() is recommended, which has the same parameter full_matrices as mindspore.ops.svd.

Categories	Subcategories	PyTorch	MindSpore	Differences
Parameters	Parameter 1	input	input	Consistent
	Parameter 2	some	full_matrices	To return the reduced singular value decomposition, MindSpore should set `full_matrices` to False, and PyTorch should set `some` to True
	Parameter 3	compute_uv	compute_uv	If `compute_uv` is False, MindSpore has only one output value s, and PyTorch has three output values u, s, v, where the values of u and v are zero-filled matrices. If `compute_uv` is True, the order of MindSpore's output values is s, u, v, and the order of PyTorch's output values is u, s, v.
	Parameter 4	out	-	For details, see General Difference Parameter Table

Code Example 1

When compute_uv is False, PyTorch has three output values.

# PyTorch
import torch
input = torch.tensor([[1, 2], [-4, -5], [2, 1]], dtype=torch.float32)
u, s, v = torch.svd(input, some=False, compute_uv=False)
print(s)
# tensor([7.0653, 1.0401])
print(u)
# tensor([[0., 0., 0.],
#         [0., 0., 0.],
#         [0., 0., 0.]])
print(v)
# tensor([[0., 0.],
#         [0., 0.]])

# MindSpore doesn't support this feature currently.

Code Example 2

When compute_uv is True, the order of output values is inconsistent. The output values of singular value decomposition are not unique.

# PyTorch
import torch
input = torch.tensor([[1, 2], [-4, -5], [2, 1]], dtype=torch.float32)
u, s, v = torch.svd(input, some=False, compute_uv=True)
print(s)
# tensor([7.0653, 1.0401])
print(u)
# tensor([[-0.3082, -0.4882,  0.8165],
#         [ 0.9061,  0.1107,  0.4082],
#         [-0.2897,  0.8657,  0.4082]])
print(v)
# tensor([[-0.6386,  0.7695],
#         [-0.7695, -0.6386]])

# MindSpore
import mindspore as ms
input = ms.Tensor([[1, 2], [-4, -5], [2, 1]], ms.float32)
s, u, v = ms.ops.svd(input, full_matrices=True, compute_uv=True)
print(s)
# [7.0652843 1.040081 ]
print(u)
# [[ 0.30821905 -0.48819482  0.81649697]
#  [-0.90613353  0.11070572  0.40824813]
#  [ 0.2896955   0.8656849   0.4082479 ]]
print(v)
# [[ 0.63863593  0.769509  ]
#  [ 0.769509   -0.63863593]]