[人工智能]torch中x = x.view(... -1)的解释

torch中在分类器中，经常会遇到：

x = x.view(x.size(0),-1)

其实在torch里面，view函数就相当于numpy的reshape，执行的操作就是对tensor进行维度转换。

import torch
x = torch.rand(2, 3, 4, 5) # 随机生成4 * 3 * 17 * 17 的数组
print("x.shape ==",x.shape) # 输出x的维度
print(x)

# 输出维度
x.shape ==torch.Size([2, 3, 4, 5])
# 生成的x
tensor([[[[0.8908, 0.6414, 0.0420, 0.7996, 0.9227],
          [0.2762, 0.7432, 0.3551, 0.3446, 0.8214],
          [0.1160, 0.8383, 0.3293, 0.7381, 0.6932],
          [0.0908, 0.6651, 0.0795, 0.7784, 0.3064]],

         [[0.7468, 0.7852, 0.7840, 0.8366, 0.4010],
          [0.6459, 0.7750, 0.9180, 0.6683, 0.2039],
          [0.5624, 0.7804, 0.1549, 0.4298, 0.1371],
          [0.0257, 0.4955, 0.4415, 0.9154, 0.4475]],

         [[0.3002, 0.7349, 0.3610, 0.6736, 0.8499],
          [0.9952, 0.8350, 0.6259, 0.7300, 0.6564],
          [0.1684, 0.1463, 0.8989, 0.9147, 0.5497],
          [0.6582, 0.9883, 0.6334, 0.7694, 0.0705]]],


        [[[0.7365, 0.5310, 0.1505, 0.2687, 0.0552],
          [0.2524, 0.9815, 0.8873, 0.9709, 0.1493],
          [0.3636, 0.2131, 0.1404, 0.9743, 0.6833],
          [0.2626, 0.3598, 0.7807, 0.0799, 0.3007]],

         [[0.7043, 0.9580, 0.7539, 0.6319, 0.0233],
          [0.7431, 0.5125, 0.5095, 0.9650, 0.5780],
          [0.8587, 0.1340, 0.9132, 0.9606, 0.3745],
          [0.9142, 0.6857, 0.0358, 0.1492, 0.3507]],

         [[0.1944, 0.9274, 0.3656, 0.0507, 0.7060],
          [0.3474, 0.9570, 0.1847, 0.0829, 0.0384],
          [0.5568, 0.5580, 0.8348, 0.4870, 0.2888],
          [0.1907, 0.5136, 0.1139, 0.0645, 0.5830]]]])

x.size(0)
2

执行view函数

x = x.view(x.size(0),-1)
x


——————————————————————————————————————output————————————————————————————————————————
tensor([[0.8908, 0.6414, 0.0420, 0.7996, 0.9227, 0.2762, 0.7432, 0.3551, 0.3446,
         0.8214, 0.1160, 0.8383, 0.3293, 0.7381, 0.6932, 0.0908, 0.6651, 0.0795,
         0.7784, 0.3064, 0.7468, 0.7852, 0.7840, 0.8366, 0.4010, 0.6459, 0.7750,
         0.9180, 0.6683, 0.2039, 0.5624, 0.7804, 0.1549, 0.4298, 0.1371, 0.0257,
         0.4955, 0.4415, 0.9154, 0.4475, 0.3002, 0.7349, 0.3610, 0.6736, 0.8499,
         0.9952, 0.8350, 0.6259, 0.7300, 0.6564, 0.1684, 0.1463, 0.8989, 0.9147,
         0.5497, 0.6582, 0.9883, 0.6334, 0.7694, 0.0705],
        [0.7365, 0.5310, 0.1505, 0.2687, 0.0552, 0.2524, 0.9815, 0.8873, 0.9709,
         0.1493, 0.3636, 0.2131, 0.1404, 0.9743, 0.6833, 0.2626, 0.3598, 0.7807,
         0.0799, 0.3007, 0.7043, 0.9580, 0.7539, 0.6319, 0.0233, 0.7431, 0.5125,
         0.5095, 0.9650, 0.5780, 0.8587, 0.1340, 0.9132, 0.9606, 0.3745, 0.9142,
         0.6857, 0.0358, 0.1492, 0.3507, 0.1944, 0.9274, 0.3656, 0.0507, 0.7060,
         0.3474, 0.9570, 0.1847, 0.0829, 0.0384, 0.5568, 0.5580, 0.8348, 0.4870,
         0.2888, 0.1907, 0.5136, 0.1139, 0.0645, 0.5830]])

x.shape == torch.Size([2, 60])

view(xxx,-1)
即，这边的这里-1表示一个不确定的数，就是你如果不确定你想要reshape成几行，但是你很肯定要reshape成xxx列。

上述例子中，

x = x.view(x.size(0),-1) 等价于  x = x.view(x.size(0),60)