Pytorch中的view()和reshape()有何不同？

Pytorch中的view()和reshape()的功能都是reshape tensor:

import torch
x = torch.arange(10)

x_2x5 = x.view(2, 5)
print(x_2x5)
x_5x2 = x.reshape(5, 2)
print(x_5x2)

其區別是：

• view()要求tensor必須是Contiguous Memory，遇到noncontiguous memory會報錯！
• reshape()沒有上述要求，在操作Contiguous Memory時，性能比view()稍差

import torch
x = torch.arange(10)
# contiguous memory
x_2x5 = x.view(2, 5)
print(x_2x5)
x_5x2 = x.reshape(5, 2)
print(x_5x2)

# noncontiguous memory
y = x_2x5.t()
y_1x10 = y.view(10)

報錯信息：

y_1x10 = y.view(10)
RuntimeError: view size is not compatible with input tensor's size and stride (at least one dimension spans across two contiguous subspaces). Use .reshape(...) instead.

解決方式：用Tensor.contiguous(memory_format=torch.contiguous_format) → Tensor
方法，將noncontiguous memory變成contiguous memory，然後再用view()

import torch
x = torch.arange(10)
# contiguous memory
x_2x5 = x.view(2, 5)
print(x_2x5)
x_5x2 = x.reshape(5, 2)
print(x_5x2)

# noncontiguous memory
y = x_2x5.t()
y_1x10 = y.contiguous().view(10)
print(y_1x10.shape)

執行結果：

torch.Size([10])