torch.Storage

torch.Storge
官方解释：A torch.Storage is a contiguous, one-dimensional array of a single data type.
Every torch.Tensor has a corresponding storage of the same data type.
有道君：torch.Storage 是一个连续的，一维的，单一数据类型的数组。每一个torch.Tensor有一个对应的torch.storage,并且二者都有相同的数据类型。

torch.Storge与torch.Tensor的区别

a = torch.FloatTensor([1, 2, 3])
b = torch.FloatStorage([1, 2, 3])

print (a)
print (type(a))
print (type(a[0]))
print (a.shape)

print (b)
print (type(b))
print (type(b[0]))
#print (b.shape) 报错

输出内容：

tensor([1., 2., 3.])
<class 'torch.Tensor'>
<class 'torch.Tensor'>
torch.Size([3])
 1.0
 2.0
 3.0
[torch.FloatStorage of size 3]
<class 'torch.FloatStorage'>
<class 'float'>

FloatTensor可以接受FloatStorage类型进行初始化，若对FloatStorage进行修改时，FloatTensor也会被修改:

b = torch.FloatStorage([1, 2, 3])
a = torch.FloatTensor(b)

b[0] = 10
print(b)
print(a)

输出内容：

 10.0
 2.0
 3.0
[torch.FloatStorage of size 3]
tensor([10.,  2.,  3.])

torch.from_numpy
torch.from_numpy与用FloatStorage初始化类似，不是直接复制数据初始化。修改numpy类型数据时，tensor数据会发生改变.

a = np.array([1,2,3])
b = torch.from_numpy(a)
a[0] = 10
print (b)

输出内容：

tensor([10,  2,  3], dtype=torch.int32)

结论
当张量需要被多处共享使用时，使用FloatStorage初始化FloatTensor.
torch.from_numpy和torch.tensor(FloatStorage)都是浅拷贝。