神經網絡模型中-model/model.name_module()的區別

標題 model / model.name_module()的區別：

對於已經訓練好的網絡模型，爲了方便後續使用和提取某一層的參數，在儲存模型時，我們會對網絡每一層進行命名+儲存參數數據。所以我們可以對每一層進行操作，例如剪枝操作。
在判斷當前層時，使用：

for name, m0 in model.name_module():
	if isinstance(m0, Conv):
		....(argument you need)
		

那麼到底print（model）和print（model.name_module()）到底有什麼區別呢？
當我們使用 print(model)時，輸出model中全部結構和層，例如：
[下列打印的輸出是博主在進行shuffleNetV2剪枝的時候輸出的模型結構]

ShuffleNetV2_2(
  (conv1): Conv2d(3, 24, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1))
  (stage2): Sequential(
    (ShuffleUnit_Stage2_0): ShuffleUnit(
      (g_conv_1x1_compress): Sequential(
        (conv1x1): Conv2d(24, 50, kernel_size=(1, 1), stride=(1, 1))
        (batch_norm): BatchNorm2d(50, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
        (relu): ReLU()
      )
      (depthwise_conv3x3): Conv2d(50, 50, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=50)
      (bn_after_depthwise): BatchNorm2d(50, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      (g_conv_1x1_expand): Sequential(
        (conv1x1): Conv2d(50, 176, kernel_size=(1, 1), stride=(1, 1), groups=2)
        (batch_norm): BatchNorm2d(176, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      )
    )
    (ShuffleUnit_Stage2_1): ShuffleUnit(
      (g_conv_1x1_compress): Sequential(
        (conv1x1): Conv2d(200, 50, kernel_size=(1, 1), stride=(1, 1), groups=2)
        (batch_norm): BatchNorm2d(50, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
        (relu): ReLU()
      )
      (depthwise_conv3x3): Conv2d(50, 50, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=50)
      (bn_after_depthwise): BatchNorm2d(50, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      (g_conv_1x1_expand): Sequential(
        (conv1x1): Conv2d(50, 200, kernel_size=(1, 1), stride=(1, 1), groups=2)
        (batch_norm): BatchNorm2d(200, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      )
    )
    (ShuffleUnit_Stage2_2): ShuffleUnit(
      (g_conv_1x1_compress): Sequential(
        (conv1x1): Conv2d(200, 50, kernel_size=(1, 1), stride=(1, 1), groups=2)
        (batch_norm): BatchNorm2d(50, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
        (relu): ReLU()
      )
      (depthwise_conv3x3): Conv2d(50, 50, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=50)
      (bn_after_depthwise): BatchNorm2d(50, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      (g_conv_1x1_expand): Sequential(
        (conv1x1): Conv2d(50, 200, kernel_size=(1, 1), stride=(1, 1), groups=2)
        (batch_norm): BatchNorm2d(200, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      )
    )
    (ShuffleUnit_Stage2_3): ShuffleUnit(
      (g_conv_1x1_compress): Sequential(
        (conv1x1): Conv2d(200, 50, kernel_size=(1, 1), stride=(1, 1), groups=2)
        (batch_norm): BatchNorm2d(50, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
        (relu): ReLU()
      )
      (depthwise_conv3x3): Conv2d(50, 50, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=50)
      (bn_after_depthwise): BatchNorm2d(50, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      (g_conv_1x1_expand): Sequential(
        (conv1x1): Conv2d(50, 200, kernel_size=(1, 1), stride=(1, 1), groups=2)
        (batch_norm): BatchNorm2d(200, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      )
    )
  )
 ])

但我們打印print(model.name_module()):

爲了方便各位觀看，這裏只截取一小部分，進行講解展示

1. model.name_module()先將model全部內容打印一邊，如同前面print（model）結果一樣
2. 打印一邊全部內容後，進行迭代，將結構中的全部block中的每一層打印一邊：
   1）先打印當前block，2）後打印當前block中的每一層
   3）以此類推，把全部結構打印出來
   舉個本文代碼的例子，結構共16個block，但最後打印出來166個結果，具體打印出來多少，根據每一個block中的層數而定！

0  ShuffleNetV2_2(
  (conv1): Conv2d(3, 24, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1), bias=False)
  (stage2): Sequential(
    (ShuffleUnit_Stage2_0): ShuffleUnit(
      (g_conv_1x1_compress): Sequential(
        (conv1x1): Conv2d(24, 50, kernel_size=(1, 1), stride=(1, 1), bias=False)
        (batch_norm): BatchNorm2d(50, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
        (relu): ReLU()
      )
      (depthwise_conv3x3): Conv2d(50, 50, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=50, bias=False)
      (bn_after_depthwise): BatchNorm2d(50, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      (g_conv_1x1_expand): Sequential(
        (conv1x1): Conv2d(50, 176, kernel_size=(1, 1), stride=(1, 1), groups=2, bias=False)
        (batch_norm): BatchNorm2d(176, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      )
    )
    (ShuffleUnit_Stage2_1): ShuffleUnit(
      (g_conv_1x1_compress): Sequential(
        (conv1x1): Conv2d(200, 50, kernel_size=(1, 1), stride=(1, 1), groups=2, bias=False)
        (batch_norm): BatchNorm2d(50, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
        (relu): ReLU()
      )
      (depthwise_conv3x3): Conv2d(50, 50, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=50, bias=False)
      (bn_after_depthwise): BatchNorm2d(50, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      (g_conv_1x1_expand): Sequential(
        (conv1x1): Conv2d(50, 200, kernel_size=(1, 1), stride=(1, 1), groups=2, bias=False)
        (batch_norm): BatchNorm2d(200, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      )
    )
    (ShuffleUnit_Stage2_2): ShuffleUnit(
      (g_conv_1x1_compress): Sequential(
        (conv1x1): Conv2d(200, 50, kernel_size=(1, 1), stride=(1, 1), groups=2, bias=False)
        (batch_norm): BatchNorm2d(50, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
        (relu): ReLU()
      )
      (depthwise_conv3x3): Conv2d(50, 50, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=50, bias=False)
      (bn_after_depthwise): BatchNorm2d(50, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      (g_conv_1x1_expand): Sequential(
        (conv1x1): Conv2d(50, 200, kernel_size=(1, 1), stride=(1, 1), groups=2, bias=False)
        (batch_norm): BatchNorm2d(200, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      )
    )
    (ShuffleUnit_Stage2_3): ShuffleUnit(
      (g_conv_1x1_compress): Sequential(
        (conv1x1): Conv2d(200, 50, kernel_size=(1, 1), stride=(1, 1), groups=2, bias=False)
        (batch_norm): BatchNorm2d(50, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
        (relu): ReLU()
      )
      (depthwise_conv3x3): Conv2d(50, 50, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=50, bias=False)
      (bn_after_depthwise): BatchNorm2d(50, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      (g_conv_1x1_expand): Sequential(
        (conv1x1): Conv2d(50, 200, kernel_size=(1, 1), stride=(1, 1), groups=2, bias=False)
        (batch_norm): BatchNorm2d(200, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      )
    )
  )

下面的輸出結果解釋什麼叫迭代輸出：

4 是當前block，內含conv，bn，relu三層
5 是迭代4中的block內容，依次輸出 5 conv，6 bn，7 relu

4 stage2.ShuffleUnit_Stage2_0.g_conv_1x1_compress Sequential(
  (conv1x1): Conv2d(24, 50, kernel_size=(1, 1), stride=(1, 1), bias=False)
  (batch_norm): BatchNorm2d(50, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
  (relu): ReLU()
)
5 stage2.ShuffleUnit_Stage2_0.g_conv_1x1_compress.conv1x1 Conv2d(24, 50, kernel_size=(1, 1), stride=(1, 1), bias=False)
6 stage2.ShuffleUnit_Stage2_0.g_conv_1x1_compress.batch_norm BatchNorm2d(50, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
7 stage2.ShuffleUnit_Stage2_0.g_conv_1x1_compress.relu ReLU()

所以在使用model中block內容的時候，一定弄清兩者區別
尤其要注意後者的輸出內容和輸出順序！！

祝好