MMLAB系列：MMCLS训练结果验证与测试

1.推理

运行demo/image_demo文件，指定模型参数运行

2. 测试

指定模型参数，运行tools/test.py文件

 3.模型修改方式

        如果需要自己写相关模块，将代码写好后，添加至相关初始化文件，然后修改模型配置文件，在相应部分修改为自己的类别即可。

        总之，模型文件以配置文件为核心

4.数据增强模型可视化展示 

        tools/visualizations/vis_pipeline.py是数据增强模型的可视化展示的脚本

参数说明：

config:模型配置文件
    --skip-type:可以跳过的步骤
    --output-dir:输出文件路径
    --phase：训练集、测试集或验证集
    --number：图片数量
    --mode：transformed表示展示最终转化后的图像，pipeline表示展示模型中间结果，concat表示展示输入和输出

设置参数并运行：

 5.MMCLS可视化模块

              tools/visualizations/vis_cam.py.py是数据增强模型的可视化展示的脚本，可以展示出某一层模型做出决策的关注点。

        使用时需要下载第三方包：pip install grad-cam

    重要参数：    

img:img图像

config：模型配置文件

checkpoint：权重文件

--target-layers：查看哪些层模型的关注点

--preview-model：打印模型

--method：可视化方法

--target-category：指定类别

修改运行配置

值得注意的是，--save-path输出路径为图片，不是文件夹

 结果如下：

6.其他工具 

        tools/analysis_tools/analyze_logs.py可以根据日志文件信息绘制图像

修改模型配置文件：

 

结果展示：

 

tools/analysis_tools/get_flops.py可以计算模型参数与计算量，同时还可以计算每一层的参数和计算量。

修改模型配置文件：

 运行结果：

ImageClassifier(
  11.229 M, 100.000% Params, 1.822 GFLOPs, 100.000% FLOPs, 
  (backbone): ResNet(
    11.177 M, 99.534% Params, 1.822 GFLOPs, 99.999% FLOPs, 
    (conv1): Conv2d(0.009 M, 0.084% Params, 0.118 GFLOPs, 6.478% FLOPs, 3, 64, kernel_size=(7, 7), stride=(2, 2), padding=(3, 3), bias=False)
    (bn1): BatchNorm2d(0.0 M, 0.001% Params, 0.002 GFLOPs, 0.088% FLOPs, 64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
    (relu): ReLU(0.0 M, 0.000% Params, 0.001 GFLOPs, 0.044% FLOPs, inplace=True)
    (maxpool): MaxPool2d(0.0 M, 0.000% Params, 0.001 GFLOPs, 0.044% FLOPs, kernel_size=3, stride=2, padding=1, dilation=1, ceil_mode=False)
    (layer1): ResLayer(
      0.148 M, 1.318% Params, 0.465 GFLOPs, 25.517% FLOPs, 
      (0): BasicBlock(
        0.074 M, 0.659% Params, 0.232 GFLOPs, 12.758% FLOPs, 
        (conv1): Conv2d(0.037 M, 0.328% Params, 0.116 GFLOPs, 6.346% FLOPs, 64, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
        (bn1): BatchNorm2d(0.0 M, 0.001% Params, 0.0 GFLOPs, 0.022% FLOPs, 64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
        (conv2): Conv2d(0.037 M, 0.328% Params, 0.116 GFLOPs, 6.346% FLOPs, 64, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
        (bn2): BatchNorm2d(0.0 M, 0.001% Params, 0.0 GFLOPs, 0.022% FLOPs, 64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
        (relu): ReLU(0.0 M, 0.000% Params, 0.0 GFLOPs, 0.022% FLOPs, inplace=True)
        (drop_path): Identity(0.0 M, 0.000% Params, 0.0 GFLOPs, 0.000% FLOPs, )
      )
      (1): BasicBlock(
        0.074 M, 0.659% Params, 0.232 GFLOPs, 12.758% FLOPs, 
        (conv1): Conv2d(0.037 M, 0.328% Params, 0.116 GFLOPs, 6.346% FLOPs, 64, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
        (bn1): BatchNorm2d(0.0 M, 0.001% Params, 0.0 GFLOPs, 0.022% FLOPs, 64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
        (conv2): Conv2d(0.037 M, 0.328% Params, 0.116 GFLOPs, 6.346% FLOPs, 64, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
        (bn2): BatchNorm2d(0.0 M, 0.001% Params, 0.0 GFLOPs, 0.022% FLOPs, 64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
        (relu): ReLU(0.0 M, 0.000% Params, 0.0 GFLOPs, 0.022% FLOPs, inplace=True)
        (drop_path): Identity(0.0 M, 0.000% Params, 0.0 GFLOPs, 0.000% FLOPs, )
      )
    )
    (layer2): ResLayer(
      0.526 M, 4.681% Params, 0.412 GFLOPs, 22.641% FLOPs, 
      (0): BasicBlock(
        0.23 M, 2.050% Params, 0.181 GFLOPs, 9.916% FLOPs, 
        (conv1): Conv2d(0.074 M, 0.657% Params, 0.058 GFLOPs, 3.173% FLOPs, 64, 128, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1), bias=False)
        (bn1): BatchNorm2d(0.0 M, 0.002% Params, 0.0 GFLOPs, 0.011% FLOPs, 128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
        (conv2): Conv2d(0.147 M, 1.313% Params, 0.116 GFLOPs, 6.346% FLOPs, 128, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
        (bn2): BatchNorm2d(0.0 M, 0.002% Params, 0.0 GFLOPs, 0.011% FLOPs, 128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
        (relu): ReLU(0.0 M, 0.000% Params, 0.0 GFLOPs, 0.011% FLOPs, inplace=True)
        (downsample): Sequential(
          0.008 M, 0.075% Params, 0.007 GFLOPs, 0.364% FLOPs, 
          (0): Conv2d(0.008 M, 0.073% Params, 0.006 GFLOPs, 0.353% FLOPs, 64, 128, kernel_size=(1, 1), stride=(2, 2), bias=False)
          (1): BatchNorm2d(0.0 M, 0.002% Params, 0.0 GFLOPs, 0.011% FLOPs, 128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
        )
        (drop_path): Identity(0.0 M, 0.000% Params, 0.0 GFLOPs, 0.000% FLOPs, )
      )
      (1): BasicBlock(
        0.295 M, 2.631% Params, 0.232 GFLOPs, 12.725% FLOPs, 
        (conv1): Conv2d(0.147 M, 1.313% Params, 0.116 GFLOPs, 6.346% FLOPs, 128, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
        (bn1): BatchNorm2d(0.0 M, 0.002% Params, 0.0 GFLOPs, 0.011% FLOPs, 128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
        (conv2): Conv2d(0.147 M, 1.313% Params, 0.116 GFLOPs, 6.346% FLOPs, 128, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
        (bn2): BatchNorm2d(0.0 M, 0.002% Params, 0.0 GFLOPs, 0.011% FLOPs, 128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
        (relu): ReLU(0.0 M, 0.000% Params, 0.0 GFLOPs, 0.011% FLOPs, inplace=True)
        (drop_path): Identity(0.0 M, 0.000% Params, 0.0 GFLOPs, 0.000% FLOPs, )
      )
    )
    (layer3): ResLayer(
      2.1 M, 18.699% Params, 0.412 GFLOPs, 22.603% FLOPs, 
      (0): BasicBlock(
        0.919 M, 8.185% Params, 0.18 GFLOPs, 9.894% FLOPs, 
        (conv1): Conv2d(0.295 M, 2.626% Params, 0.058 GFLOPs, 3.173% FLOPs, 128, 256, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1), bias=False)
        (bn1): BatchNorm2d(0.001 M, 0.005% Params, 0.0 GFLOPs, 0.006% FLOPs, 256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
        (conv2): Conv2d(0.59 M, 5.253% Params, 0.116 GFLOPs, 6.346% FLOPs, 256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
        (bn2): BatchNorm2d(0.001 M, 0.005% Params, 0.0 GFLOPs, 0.006% FLOPs, 256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
        (relu): ReLU(0.0 M, 0.000% Params, 0.0 GFLOPs, 0.006% FLOPs, inplace=True)
        (downsample): Sequential(
          0.033 M, 0.296% Params, 0.007 GFLOPs, 0.358% FLOPs, 
          (0): Conv2d(0.033 M, 0.292% Params, 0.006 GFLOPs, 0.353% FLOPs, 128, 256, kernel_size=(1, 1), stride=(2, 2), bias=False)
          (1): BatchNorm2d(0.001 M, 0.005% Params, 0.0 GFLOPs, 0.006% FLOPs, 256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
        )
        (drop_path): Identity(0.0 M, 0.000% Params, 0.0 GFLOPs, 0.000% FLOPs, )
      )
      (1): BasicBlock(
        1.181 M, 10.515% Params, 0.232 GFLOPs, 12.709% FLOPs, 
        (conv1): Conv2d(0.59 M, 5.253% Params, 0.116 GFLOPs, 6.346% FLOPs, 256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
        (bn1): BatchNorm2d(0.001 M, 0.005% Params, 0.0 GFLOPs, 0.006% FLOPs, 256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
        (conv2): Conv2d(0.59 M, 5.253% Params, 0.116 GFLOPs, 6.346% FLOPs, 256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
        (bn2): BatchNorm2d(0.001 M, 0.005% Params, 0.0 GFLOPs, 0.006% FLOPs, 256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
        (relu): ReLU(0.0 M, 0.000% Params, 0.0 GFLOPs, 0.006% FLOPs, inplace=True)
        (drop_path): Identity(0.0 M, 0.000% Params, 0.0 GFLOPs, 0.000% FLOPs, )
      )
    )
    (layer4): ResLayer(
      8.394 M, 74.752% Params, 0.411 GFLOPs, 22.583% FLOPs, 
      (0): BasicBlock(
        3.673 M, 32.711% Params, 0.18 GFLOPs, 9.883% FLOPs, 
        (conv1): Conv2d(1.18 M, 10.506% Params, 0.058 GFLOPs, 3.173% FLOPs, 256, 512, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1), bias=False)
        (bn1): BatchNorm2d(0.001 M, 0.009% Params, 0.0 GFLOPs, 0.003% FLOPs, 512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
        (conv2): Conv2d(2.359 M, 21.011% Params, 0.116 GFLOPs, 6.346% FLOPs, 512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
        (bn2): BatchNorm2d(0.001 M, 0.009% Params, 0.0 GFLOPs, 0.003% FLOPs, 512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
        (relu): ReLU(0.0 M, 0.000% Params, 0.0 GFLOPs, 0.003% FLOPs, inplace=True)
        (downsample): Sequential(
          0.132 M, 1.176% Params, 0.006 GFLOPs, 0.355% FLOPs, 
          (0): Conv2d(0.131 M, 1.167% Params, 0.006 GFLOPs, 0.353% FLOPs, 256, 512, kernel_size=(1, 1), stride=(2, 2), bias=False)
          (1): BatchNorm2d(0.001 M, 0.009% Params, 0.0 GFLOPs, 0.003% FLOPs, 512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
        )
        (drop_path): Identity(0.0 M, 0.000% Params, 0.0 GFLOPs, 0.000% FLOPs, )
      )
      (1): BasicBlock(
        4.721 M, 42.040% Params, 0.231 GFLOPs, 12.701% FLOPs, 
        (conv1): Conv2d(2.359 M, 21.011% Params, 0.116 GFLOPs, 6.346% FLOPs, 512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
        (bn1): BatchNorm2d(0.001 M, 0.009% Params, 0.0 GFLOPs, 0.003% FLOPs, 512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
        (conv2): Conv2d(2.359 M, 21.011% Params, 0.116 GFLOPs, 6.346% FLOPs, 512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
        (bn2): BatchNorm2d(0.001 M, 0.009% Params, 0.0 GFLOPs, 0.003% FLOPs, 512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
        (relu): ReLU(0.0 M, 0.000% Params, 0.0 GFLOPs, 0.003% FLOPs, inplace=True)
        (drop_path): Identity(0.0 M, 0.000% Params, 0.0 GFLOPs, 0.000% FLOPs, )
      )
    )
  )
  init_cfg=[{'type': 'Kaiming', 'layer': ['Conv2d']}, {'type': 'Constant', 'val': 1, 'layer': ['_BatchNorm', 'GroupNorm']}]
  (neck): GlobalAveragePooling(
    0.0 M, 0.000% Params, 0.0 GFLOPs, 0.001% FLOPs, 
    (gap): AdaptiveAvgPool2d(0.0 M, 0.000% Params, 0.0 GFLOPs, 0.001% FLOPs, output_size=(1, 1))
  )
  (head): LinearClsHead(
    0.052 M, 0.466% Params, 0.0 GFLOPs, 0.000% FLOPs, 
    (compute_loss): CrossEntropyLoss(0.0 M, 0.000% Params, 0.0 GFLOPs, 0.000% FLOPs, )
    (compute_accuracy): Accuracy(0.0 M, 0.000% Params, 0.0 GFLOPs, 0.000% FLOPs, )
    (fc): Linear(0.052 M, 0.466% Params, 0.0 GFLOPs, 0.000% FLOPs, in_features=512, out_features=102, bias=True)
  )
  init_cfg={'type': 'Normal', 'layer': 'Linear', 'std': 0.01}
)
==============================
Input shape: (3, 224, 224)
Flops: 1.82 GFLOPs
Params: 11.23 M
==============================