比较与torch.sparse_coo_tensor的功能差异
torch.sparse_coo_tensor
torch.sparse_coo_tensor(
indices,
values,
size=None,
dtype=None,
device=None,
requires_grad=False
)
更多内容详见torch.sparse_coo_tensor。
mindspore.SparseTensor
class mindspore.SparseTensor(
indices,
values,
dense_shape
)
更多内容详见mindspore.SparseTensor。
使用方式
PyTorch:以COO(rdinate)
格式构造一个稀疏张量。
MindSpore:构造稀疏张量,只能在Cell
的构造方法中使用,PyNative模式暂不支持稀疏张量。
代码示例
# 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)