ONNXRuntime学习笔记(二)

继上一篇计划的实践项目，这篇记录我训练模型相关的工作。

• 首先要确定总体目标：训练一个pytorch模型，CIFAR-100数据集测试集acc达到90%；部署后推理效率达到50ms/张, 部署平台为window10+3050Ti+RX5800h.
• 训练模型的话，最好是有一套完备的代码，像谷歌的models，FB的detectron2，商汤的mm系列等等框架，这些是建立在深度学习框架tf或pth基础上的进一步封装，提供一些更高级的写好的模块可以调用，如Resnet、FPN、、proposal、NMS等等。但凡事都有两面，封装度越高意味着稳定性更好但修改的灵活性越差。只调用API对我们理解底层实现是不利的。之前我写过一个基于Pytorch的图像分类训练推理代码，现在又可以拿出来用一用了，地址：https://github.com/lee-zq/CNN-Backbone ，我在之前训练CIFAR-10的基础上又添加了CIFAR-100数据集的Dataloader创建代码。

1. 首先，我尝试了CIFAR10+DenseNet，最后测试效果Acc=85%；然后尝试了CIFAR10+ResNet18，收敛较慢，但最终Acc=91.02%；基于此，。我尝试了CIFAR100+ResNet18，收敛很慢，大概到73Epoch稳定下来，但最终训练集Acc能达到90.62%，但测试集Acc为65.67%。大概率原因是模型拟合能力够用但是训练集多样性太差。模型结构如下：

ResNet(
  (conv1): Sequential(
    (0): Conv2d(3, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
    (1): BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
    (2): ReLU()
  )
  (layer1): Sequential(
    (0): ResidualBlock(
      (left): Sequential(
        (0): Conv2d(64, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
        (1): BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
        (2): ReLU(inplace=True)
        (3): Conv2d(64, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
        (4): BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      )
      (shortcut): Sequential()
    )
    (1): ResidualBlock(
      (left): Sequential(
        (0): Conv2d(64, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
        (1): BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
        (2): ReLU(inplace=True)
        (3): Conv2d(64, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
        (4): BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      )
      (shortcut): Sequential()
    )
  )
  (layer2): Sequential(
    (0): ResidualBlock(
      (left): Sequential(
        (0): Conv2d(64, 128, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1), bias=False)
        (1): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
        (2): ReLU(inplace=True)
        (3): Conv2d(128, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
        (4): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      )
      (shortcut): Sequential(
        (0): Conv2d(64, 128, kernel_size=(1, 1), stride=(2, 2), bias=False)
        (1): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      )
    )
    (1): ResidualBlock(
      (left): Sequential(
        (0): Conv2d(128, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
        (1): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
        (2): ReLU(inplace=True)
        (3): Conv2d(128, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
        (4): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      )
      (shortcut): Sequential()
    )
  )
  (layer3): Sequential(
    (0): ResidualBlock(
      (left): Sequential(
        (0): Conv2d(128, 256, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1), bias=False)
        (1): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
        (2): ReLU(inplace=True)
        (3): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
        (4): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      )
      (shortcut): Sequential(
        (0): Conv2d(128, 256, kernel_size=(1, 1), stride=(2, 2), bias=False)
        (1): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      )
    )
    (1): ResidualBlock(
      (left): Sequential(
        (0): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
        (1): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
        (2): ReLU(inplace=True)
        (3): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
        (4): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      )
      (shortcut): Sequential()
    )
  )
  (layer4): Sequential(
    (0): ResidualBlock(
      (left): Sequential(
        (0): Conv2d(256, 512, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1), bias=False)
        (1): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
        (2): ReLU(inplace=True)
        (3): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
        (4): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      )
      (shortcut): Sequential(
        (0): Conv2d(256, 512, kernel_size=(1, 1), stride=(2, 2), bias=False)
        (1): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      )
    )
    (1): ResidualBlock(
      (left): Sequential(
        (0): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
        (1): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
        (2): ReLU(inplace=True)
        (3): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
        (4): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      )
      (shortcut): Sequential()
    )
  )
  (fc): Linear(in_features=512, out_features=10, bias=True)
)
Total number of parameters: 11173962

总参数量约11M，既然CIFAR-100效果太差，那就暂且还是用CIFAR-10做后面的训练测试吧，我又在之前的数据增强基础上加了RandomGrayscale和RandomAffine，最终的数据增强如下：

        self.mean = [0.4914, 0.4822, 0.4465]
        self.std = [0.2023, 0.1994, 0.2010]
        self.num_workers= num_workers
        self.transform_train = transforms.Compose([# 数据增强
                                    transforms.RandomCrop(32, padding=4),
                                    transforms.RandomHorizontalFlip(),
                                    transforms.RandomGrayscale(0.15),
                                    transforms.RandomAffine((-30,30)),
                                    transforms.RandomRotation(20),  
                                    transforms.ToTensor(),
                                    transforms.Normalize(self.mean, self.std),
                                    transforms.RandomErasing(),
        ])

2. 然后微调继续训练，测试集Acc进一步提升到92.28%，可见数据多样性的重要性。进一步的，torchvision提供了AutoAugment数据增强方法的接口，可以直接调用，最终数据增强代码如下：

        self.mean = [0.4914, 0.4822, 0.4465]
        self.std = [0.2023, 0.1994, 0.2010]
        self.num_workers= num_workers
        self.transform_train = transforms.Compose([# 数据增强
                                    transforms.RandomCrop(32, padding=4),
                                    transforms.RandomHorizontalFlip(),
                                    transforms.autoaugment.AutoAugment(policy=transforms.autoaugment.AutoAugmentPolicy.CIFAR10),
                                    transforms.ToTensor(),
                                    transforms.Normalize(self.mean, self.std),
                                    transforms.RandomErasing()
        ])

3. 训练epoch数为80，优化器Adam，初始学习率0.01，每20epoch衰减，衰减因子gamma为0.1，目前还在训练ing，要花两个小时。完整重头训练估计要花4个小时，在之前的基础上微调会快很多，最终测试集Acc达到94.83%，达到预期。下一篇记录利用onnxruntime推理进行测试的过程。