Function Differences with torch.sparse_coo_tensor

torch.sparse_coo_tensor

torch.sparse_coo_tensor(
  indices,
  values,
  size=None,
  dtype=None,
  device=None,
  requires_grad=False
)

For more information, see torch.sparse_coo_tensor.

mindspore.SparseTensor

class mindspore.SparseTensor(
  indices,
  values,
  dense_shape
)

For more information, see mindspore.SparseTensor.

Differences

PyTorch: Constructs a sparse tensors in COO(rdinate) format.

MindSpore：Constructs a sparse tensors. It can only be used in the Cell construct method. PyNative mode does not support sparse tensor.

Code Example

# In MindSpore：
import mindspore as ms
import mindspore.nn as nn

class Net(nn.Cell):
    def __init__(self, shape):
        super(Net, self).__init__()
        self.shape = shape
    def construct(self, indices, values):
        x = ms.SparseTensor(indices, values, self.shape)
        return x.indices, x.values, x.shape

indices = ms.Tensor([[0, 1], [1, 2]])
values = ms.Tensor([1, 2], dtype=ms.float32)
out = Net((3, 4))(indices, values)
print(out[0])
print(out[1])
print(out[2])
# Linux Out:
# [[0 1]
#  [1 2]]
# [1. 2.]
# (3, 4)
# Windows Out:
# [[0 1]
#  [1 2]]
# [1. 2.]
# (Tensor(shape=[], dtype=Int64, value= 3), Tensor(shape=[], dtype=Int64, value= 4))

# In torch:
import torch
i = torch.tensor([[0, 1],
                  [1, 2]])
v = torch.tensor([1, 2], dtype=torch.float32)
out = torch.sparse_coo_tensor(i, v, [3, 4])
print(out)
# Out:
# tensor(indices=tensor([[0, 1],
#                       [1, 2]]),
#        values=tensor([1., 2.]),
#        size=(3, 4), nnz=2, layout=torch.sparse_coo)