Pytorch中[:,None]的用法解析

一.Pytorch中[:,None]的用法解析

1. [:,None]的用法解析

Tensor中利用None来增加维度，可以简单的理解为在None的位置上增加一维，新增维度大小为1，同时有几个None就会增加几个维度。

2. 代码举例

2.1 [None,:,:]和[None,:]输出张量一样，输入数据为二维张量，输出为三维张量。

x = torch.randn(3,4)
y = x[None,:,:]
x,x.shape,y,y.shape

输出结果如下：
(tensor([[-1.2004,  0.0202, -0.4225, -0.0444],
         [ 0.4218, -1.0867,  1.4224,  0.5967],
         [ 0.4703,  0.8158, -0.9460, -0.7291]]),
 torch.Size([3, 4]),
 tensor([[[-1.2004,  0.0202, -0.4225, -0.0444],
          [ 0.4218, -1.0867,  1.4224,  0.5967],
          [ 0.4703,  0.8158, -0.9460, -0.7291]]]),
 torch.Size([1, 3, 4]))

2.2 [None,None:,:]和[None,None]输出张量一样，输入数据为二维张量，输出为四维张量。

x = torch.randn(3,4)
y = x[None,None,:,:]
x,x.shape,y,y.shape

输出结果如下：
(tensor([[-0.7301, -0.2588,  0.2528, -0.7637],
         [-1.5438,  0.6894,  0.5747,  0.0481],
         [ 0.5045, -0.3611, -1.2757,  1.0789]]),
 torch.Size([3, 4]),
 tensor([[[[-0.7301, -0.2588,  0.2528, -0.7637],
           [-1.5438,  0.6894,  0.5747,  0.0481],
           [ 0.5045, -0.3611, -1.2757,  1.0789]]]]),
 torch.Size([1, 1, 3, 4]))

2.3 [:,None,:]和[:,None]输出张量一样，输入数据为二维张量，输出为三维张量。

x = torch.randn(3,4)
y = x[:,None,:]
x,x.shape,y,y.shape

输出结果如下：
(tensor([[ 1.0753, -1.5525,  0.8249, -0.2986],
         [-0.7956, -0.0708,  1.8574, -1.0563],
         [ 0.5642,  0.9701,  1.0636, -1.2102]]),
 torch.Size([3, 4]),
 tensor([[[ 1.0753, -1.5525,  0.8249, -0.2986]],
 
         [[-0.7956, -0.0708,  1.8574, -1.0563]],
 
         [[ 0.5642,  0.9701,  1.0636, -1.2102]]]),
 torch.Size([3, 1, 4]))

2.4 [:,None,:]和[:,None]输出张量一样，输入数据为二维张量，输出为三维张量。

x = torch.randn(3,4)
y = x[:,None]
x,x.shape,y,y.shape

输出结果如下：
(tensor([[-2.7815, -0.8274,  1.1110,  0.9889],
         [-0.6636, -1.5992,  0.7225,  0.3466],
         [-1.4326,  2.0451, -1.6679,  0.0902]]),
 torch.Size([3, 4]),
 tensor([[[-2.7815, -0.8274,  1.1110,  0.9889]],
 
         [[-0.6636, -1.5992,  0.7225,  0.3466]],
 
         [[-1.4326,  2.0451, -1.6679,  0.0902]]]),
 torch.Size([3, 1, 4]))

2.5 [:,:,None]，输入数据为二维张量，输出为三维张量。

x = torch.randn(3,4)
y = x[:,:,None]
x,x.shape,y,y.shape

输出结果如下：
(tensor([[ 0.0735,  0.1196,  1.7420, -0.2371],
         [-0.2613, -1.3396, -0.0262, -0.3695],
         [-1.2122, -1.1700,  2.3281, -0.8234]]),
 torch.Size([3, 4]),
 tensor([[[ 0.0735],
          [ 0.1196],
          [ 1.7420],
          [-0.2371]],
 
         [[-0.2613],
          [-1.3396],
          [-0.0262],
          [-0.3695]],
 
         [[-1.2122],
          [-1.1700],
          [ 2.3281],
          [-0.8234]]]),
 torch.Size([3, 4, 1]))

2.6 [:,:,None,None]，输入数据为二维张量，输出为四维张量。

x = torch.randn(3,4)
y = x[:,:,None,None]
x,x.shape,y,y.shape

输出结果如下：
(tensor([[ 0.5205,  0.7751, -0.9279,  0.6369],
         [ 1.0077, -0.2766,  0.5953, -1.1734],
         [ 1.9789,  0.1456, -1.9392, -0.4931]]),
 torch.Size([3, 4]),
 tensor([[[[ 0.5205]],
 
          [[ 0.7751]],
 
          [[-0.9279]],
 
          [[ 0.6369]]],
 
 
         [[[ 1.0077]],
 
          [[-0.2766]],
 
          [[ 0.5953]],
 
          [[-1.1734]]],
 
 
         [[[ 1.9789]],
 
          [[ 0.1456]],
 
          [[-1.9392]],
 
          [[-0.4931]]]]),
 torch.Size([3, 4, 1, 1]))