10fromtypingimportList1112importtorch.nnasnn1314fromlabmlimportlab15fromlabml.configsimportoption16fromlabml\_helpers.datasets.cifar10importCIFAR10ConfigsasCIFAR10DatasetConfigs17fromlabml\_helpers.moduleimportModule18fromlabml\_nn.experiments.mnistimportMNISTConfigs

21classCIFAR10Configs(CIFAR10DatasetConfigs,MNISTConfigs):

30dataset\_name:str='CIFAR10'

33@option(CIFAR10Configs.train\_dataset)34defcifar10\_train\_augmented():

38fromtorchvision.datasetsimportCIFAR1039fromtorchvision.transformsimporttransforms40returnCIFAR10(str(lab.get\_data\_path()),41train=True,42download=True,43transform=transforms.Compose([

45transforms.RandomCrop(32,padding=4),

47transforms.RandomHorizontalFlip(),

49transforms.ToTensor(),50transforms.Normalize((0.5,0.5,0.5),(0.5,0.5,0.5))51]))

54@option(CIFAR10Configs.valid\_dataset)55defcifar10\_valid\_no\_augment():

59fromtorchvision.datasetsimportCIFAR1060fromtorchvision.transformsimporttransforms61returnCIFAR10(str(lab.get\_data\_path()),62train=False,63download=True,64transform=transforms.Compose([65transforms.ToTensor(),66transforms.Normalize((0.5,0.5,0.5),(0.5,0.5,0.5))67]))

70classCIFAR10VGGModel(Module):

75defconv\_block(self,in\_channels,out\_channels)-\>nn.Module:

79returnnn.Sequential(80nn.Conv2d(in\_channels,out\_channels,kernel\_size=3,padding=1),81nn.ReLU(inplace=True),82)

84def\_\_init\_\_(self,blocks:List[List[int]]):85super().\_\_init\_\_()

89assertlen(blocks)==590layers=[]

92in\_channels=3

94forblockinblocks:

96forchannelsinblock:97layers+=self.conv\_block(in\_channels,channels)98in\_channels=channels

100layers+=[nn.MaxPool2d(kernel\_size=2,stride=2)]

103self.layers=nn.Sequential(\*layers)

105self.fc=nn.Linear(in\_channels,10)

107defforward(self,x):

109x=self.layers(x)

111x=x.view(x.shape[0],-1)

113returnself.fc(x)