pytorch学习笔记——网络架构四调整与数据再增强（10）

训练集和测试集准确率纠缠不清，准确有没有再提升的可能？一种可能的情况就是过拟合了。基于之前训练准确率与测试集准确率十分接近的情况，我们再次增加网络的规模，同时使用对比度变换的数据增强方案，看看是不是有可能使准确率进一步提升。改变后的网络架构如下：

SimpleNet(
  (conv1): Sequential(
    (0): Conv2d(1, 8, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
    (1): BatchNorm2d(8, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
    (2): ReLU()
    (3): Conv2d(8, 16, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
    (4): BatchNorm2d(16, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
    (5): ReLU()
    (6): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)
  )
  (conv2): Sequential(
    (0): Conv2d(16, 32, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
    (1): BatchNorm2d(32, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
    (2): ReLU()
    (3): Conv2d(32, 48, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
    (4): BatchNorm2d(48, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
    (5): ReLU()
    (6): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)
  )
  (fc): Sequential(
    (0): Linear(in_features=192, out_features=192, bias=True)
    (1): Dropout(p=0.5, inplace=False)
    (2): Linear(in_features=192, out_features=128, bias=True)
    (3): Dropout(p=0.5, inplace=False)
    (4): Linear(in_features=128, out_features=10, bias=True)
  )
)>
----------------------------------------------------------------
        Layer (type)               Output Shape         Param #
================================================================
            Conv2d-1              [-1, 8, 8, 8]              80
       BatchNorm2d-2              [-1, 8, 8, 8]              16
              ReLU-3              [-1, 8, 8, 8]               0
            Conv2d-4             [-1, 16, 8, 8]           1,168
       BatchNorm2d-5             [-1, 16, 8, 8]              32
              ReLU-6             [-1, 16, 8, 8]               0
         MaxPool2d-7             [-1, 16, 4, 4]               0
            Conv2d-8             [-1, 32, 4, 4]           4,640
       BatchNorm2d-9             [-1, 32, 4, 4]              64
             ReLU-10             [-1, 32, 4, 4]               0
           Conv2d-11             [-1, 48, 4, 4]          13,872
      BatchNorm2d-12             [-1, 48, 4, 4]              96
             ReLU-13             [-1, 48, 4, 4]               0
        MaxPool2d-14             [-1, 48, 2, 2]               0
           Linear-15                  [-1, 192]          37,056
          Dropout-16                  [-1, 192]               0
           Linear-17                  [-1, 128]          24,704
          Dropout-18                  [-1, 128]               0
           Linear-19                   [-1, 10]           1,290
================================================================
Total params: 83,018
Trainable params: 83,018
Non-trainable params: 0
----------------------------------------------------------------
Input size (MB): 0.00
Forward/backward pass size (MB): 0.07
Params size (MB): 0.32
Estimated Total Size (MB): 0.39
----------------------------------------------------------------

得出的结果如下：

作用相当于挠痒痒。删去对比度增强，试训练一下再试试：

害，结果还不如之前的架构呢。具体原因我还得仔细想想。