CIFAR10 Experiment for Instance Normalization

This demonstrates the use of an instance normalization layer in a convolutional neural network for classification. Not that instance normalization was designed for style transfer and this is only a demo.

16importtorch.nnasnn1718fromlabmlimportexperiment19fromlabml.configsimportoption20fromlabml\_nn.experiments.cifar10importCIFAR10Configs,CIFAR10VGGModel21fromlabml\_nn.normalization.instance\_normimportInstanceNorm

VGG model for CIFAR-10 classification

This derives from the generic VGG style architecture.

24classModel(CIFAR10VGGModel):

31defconv\_block(self,in\_channels,out\_channels)-\>nn.Module:32returnnn.Sequential(33nn.Conv2d(in\_channels,out\_channels,kernel\_size=3,padding=1),34InstanceNorm(out\_channels),35nn.ReLU(inplace=True),36)

38def\_\_init\_\_(self):39super().\_\_init\_\_([[64,64],[128,128],[256,256,256],[512,512,512],[512,512,512]])

Create model

42@option(CIFAR10Configs.model)43def\_model(c:CIFAR10Configs):

47returnModel().to(c.device)

50defmain():

Create experiment

52experiment.create(name='cifar10',comment='instance norm')

Create configurations

54conf=CIFAR10Configs()

Load configurations

56experiment.configs(conf,{57'optimizer.optimizer':'Adam',58'optimizer.learning\_rate':2.5e-4,59})

Start the experiment and run the training loop

61withexperiment.start():62conf.run()

66if\_\_name\_\_=='\_\_main\_\_':67main()

labml.ai