CIFAR10 Experiment for Group Normalization

12importtorch.nnasnn1314fromlabmlimportexperiment15fromlabml.configsimportoption16fromlabml\_nn.experiments.cifar10importCIFAR10Configs,CIFAR10VGGModel17fromlabml\_nn.normalization.batch\_normimportBatchNorm

VGG model for CIFAR-10 classification

This derives from the generic VGG style architecture.

20classModel(CIFAR10VGGModel):

27defconv\_block(self,in\_channels,out\_channels)-\>nn.Module:28returnnn.Sequential(29nn.Conv2d(in\_channels,out\_channels,kernel\_size=3,padding=1),30BatchNorm(out\_channels),31nn.ReLU(inplace=True),32)

34def\_\_init\_\_(self):35super().\_\_init\_\_([[64,64],[128,128],[256,256,256],[512,512,512],[512,512,512]])

Create model

38@option(CIFAR10Configs.model)39defmodel(c:CIFAR10Configs):

43returnModel().to(c.device)

46defmain():

Create experiment

48experiment.create(name='cifar10',comment='batch norm')

Create configurations

50conf=CIFAR10Configs()

Load configurations

52experiment.configs(conf,{53'optimizer.optimizer':'Adam',54'optimizer.learning\_rate':2.5e-4,55'train\_batch\_size':64,56})

Start the experiment and run the training loop

58withexperiment.start():59conf.run()

63if\_\_name\_\_=='\_\_main\_\_':64main()

labml.ai