复制或创建零张量

Question

什么是使用torch-tensor的好方法：复制或创建新的张量？速度呢？ 目标是在每次调用方法时获得一个具有恒定大小的新零张量。

class CopyOrCreateZeros:
    def __init__(self):
        self.fx = torch.zeros((800, 600))
        pass

    def method(self, x):
        """ this method will be used many ofter e.g. in loop """
        # do something

        new_x1 = torch.copy(self.fx)
        new_x2 = torch.clone(self.fx)
        new_x3 = x.new_zeros()
        new_x4 = torch.zeros_like(x)
        new_x5 = torch.zeros()

Answer 1

似乎克隆fx 张量是最慢的解决方案，而其他三个（即Tensor.new_zeros、torch.zeros_like 和torch.zeros）的时间相当：

• torch.clone:

  %timeit fx.clone()
  1000 loops, best of 5: 277 µs per loop
  

• torch.Tensor.new_zeros:

  %timeit fx.new_zeros(fx.shape)
  1000 loops, best of 5: 257 µs per loop
  

• torch.zeros_like:

  %timeit torch.zeros_like(fx)
  1000 loops, best of 5: 258 µs per loop
  

• torch.zeros:

  %timeit torch.zeros(fx.shape)
  1000 loops, best of 5: 256 µs per loop