tensor[:,0]意思应该是取第0个channel的维度，其他维度不变

代码一：

import torch
a=torch.randn((2,1,3,4))
print('a=',a)
b=a[:,0]
print('b=',b)

a= tensor([[[[-0.3767, -0.8647, -0.6132, -0.5862],
          [ 0.5800,  0.3467, -0.4182, -0.0916],
          [ 0.6316, -0.6858, -0.2674,  0.1670]]],


        [[[-0.1980, -0.7904, -1.1460,  1.4207],
          [-0.0893,  0.2657, -0.1714,  0.0230],
          [ 0.1632,  0.8553,  1.6556, -0.1242]]]])
b= tensor([[[-0.3767, -0.8647, -0.6132, -0.5862],
         [ 0.5800,  0.3467, -0.4182, -0.0916],
         [ 0.6316, -0.6858, -0.2674,  0.1670]],

        [[-0.1980, -0.7904, -1.1460,  1.4207],
         [-0.0893,  0.2657, -0.1714,  0.0230],
         [ 0.1632,  0.8553,  1.6556, -0.1242]]])

可以看出代码一相当于把通道维度压缩了

代码二：

import torch
a=torch.randn((2,2,3,4))
print('a=',a)
b=a[:,0]
print('b=',b)

代码二模拟生成两个特征图，每个特征图有两个channel，宽高为3和4

a= tensor([[[[ 0.1237,  0.9523, -0.0208, -0.8758],
          [ 0.4308,  0.3393,  0.0697,  0.0206],
          [ 0.5322,  0.3302, -2.2941,  1.1660]],

         [[-0.2208,  0.4693, -0.8774,  1.4004],
          [-2.5176, -0.3753,  0.1951,  0.3845],
          [ 1.1252,  0.6507, -0.8537,  0.8822]]],


        [[[-0.8990,  1.0145,  1.1724, -1.3846],
          [-0.1402, -2.3992, -0.0735,  0.7676],
          [-1.5613,  0.1181,  2.2502,  1.7849]],

         [[ 0.0421,  1.0093, -0.5084,  0.8766],
          [-1.0246,  0.2210, -0.5652, -1.4182],
          [ 0.1860, -0.3156, -0.4796, -0.6364]]]])
b= tensor([[[ 0.1237,  0.9523, -0.0208, -0.8758],
         [ 0.4308,  0.3393,  0.0697,  0.0206],
         [ 0.5322,  0.3302, -2.2941,  1.1660]],

        [[-0.8990,  1.0145,  1.1724, -1.3846],
         [-0.1402, -2.3992, -0.0735,  0.7676],
         [-1.5613,  0.1181,  2.2502,  1.7849]]])

 

 红色部分为黄色部分组合而成，相当于只保留了这两个特征图的第0通道