Pytoch squeeze() unsqueeze() usage

brief introduction

torch.squeeze(input, dim=None, out=None)： Compress the dimension of data , Remove the dimension as 1 Dimension of .
squeeze The functionality ： Remove size by 1 Dimensions , Including rows and columns . When the dimension is greater than or equal to 2 when ,squeeze() No use .
squeeze(0)： If the value of the first dimension is 1 Remove the first dimension , for example a.squeeze(0),a by torch.tensor() Format tensor .
squeeze(1)： If the value of the second dimension is 1 Remove the second dimension
squeeze(-1)： Remove the final dimension value as 1 Dimensions

torch.unsqueeze (input, dim=None, out=None)： Expand the dimension of data , Upgrade dimension .

The format used can be torch.unsqueeze(x, 0), It can also be x.unsqueeze(0).

The sample code

a = torch.Tensor(1, 3)
print(a)
print(a.squeeze(0))
print(a.squeeze(1))

b = torch.Tensor(2, 3)
print(b)
print(b.squeeze(0))
print(b.squeeze(1))

c = torch.Tensor(3, 1)
print(c)
print(c.squeeze(0))
print(c.squeeze(1))

x = torch.tensor([1, 2, 3, 4])
print(x)
print(torch.unsqueeze(x, 0))
print(torch.unsqueeze(x, 1))

Process analysis

Define tensor a, by 2 dimension , The first dimension is 1 Elements , The second dimension is 3 Elements .
Output ：tensor([[2.6994e-30, 2.4164e-13, 1.8392e-13]])
adopt a.squeeze(0) Reduce the dimension of the first dimension , At this time, the first dimension has 1 Elements , Dimensionality reduction , The first dimension disappears , The second dimension automatically becomes the first dimension with three elements , And a comparison , That is, it disappears “[]”.
Output ：tensor([2.6994e-30, 2.4164e-13, 1.8392e-13])
adopt a.squeeze(1) Reduce the dimension of the second dimension , At this time, the first dimension has 3 Elements , Dimensionality cannot be reduced , Do not operate , Output and a identical .
Output ：tensor([[2.6994e-30, 2.4164e-13, 1.8392e-13]])

Define tensor b, by 2 dimension , The first dimension is 2 Elements , The second dimension is 3 Elements .
First of all 、 Two dimensions cannot be reduced , Because the three outputs are the same .
Output ：

tensor([[0., 0., 0.],
        [0., 0., 0.]])
tensor([[0., 0., 0.],
        [0., 0., 0.]])
tensor([[0., 0., 0.],
        [0., 0., 0.]])

Define tensor c, by 2 dimension , The first dimension is 3 Elements , The second dimension is 1 Elements .
Output ：

tensor([[0.0000e+00],
        [       nan],
        [5.2781e-24]])

adopt c.squeeze(0) Reduce the dimension of the first dimension , At this time, the first dimension has 3 Elements , Dimensionality cannot be reduced , Do not operate , Output and c identical .
Output ：

tensor([[0.0000e+00],
        [       nan],
        [5.2781e-24]])

adopt c.squeeze(1) Reduce the dimension of the second dimension , At this time, the second dimension has 1 Elements , Dimensionality reduction , The second dimension disappears , The value of the second dimension automatically enters the first dimension .
Output ：
tensor([0.0000e+00, nan, 5.2781e-24])

Define tensor x, by 1 dimension , The values are 1, 2, 3, 4.
Output ：tensor([1, 2, 3, 4])
adopt x.unsqueeze(0) Add a dimension to the first dimension , Change the original tensor into 2 dimension , Dimension becomes (1, 4). And x comparison , That is to add a layer “[]”.
Output ：tensor([[1, 2, 3, 4]])
adopt x.unsqueeze(1) Add a dimension to the second dimension , Change the original tensor into 2 dimension , Dimension becomes (4, 1).
Output ：tensor([[1], [2], [3], [4]])

Running results

tensor([[2.6994e-30, 2.4164e-13, 1.8392e-13]])
tensor([2.6994e-30, 2.4164e-13, 1.8392e-13])
tensor([[2.6994e-30, 2.4164e-13, 1.8392e-13]])
tensor([[0., 0., 0.],
        [0., 0., 0.]])
tensor([[0., 0., 0.],
        [0., 0., 0.]])
tensor([[0., 0., 0.],
        [0., 0., 0.]])
tensor([[0.0000e+00],
        [       nan],
        [5.2781e-24]])
tensor([[0.0000e+00],
        [       nan],
        [5.2781e-24]])
tensor([0.0000e+00,        nan, 5.2781e-24])
tensor([1, 2, 3, 4])
tensor([[1, 2, 3, 4]])
tensor([[1],
        [2],
        [3],
        [4]])