Update Training a Classifier tutorial — replace deprecated transforms, use torch.accelerator

[sekyondaMeta]

Description

The Training a Classifier tutorial contains deprecated APIs and outdated patterns that should be modernized.

Changes needed

Deprecated APIs

Issue	Current Code	Replacement	Since
transforms.ToTensor() is deprecated	transforms.ToTensor()	transforms.v2.ToImage() + transforms.v2.ToDtype(torch.float32, scale=True)	torchvision 0.16+ (v2 transforms). ToTensor() deprecated in torchvision 0.18, will be removed in a future version.
transforms.Normalize() from v1 namespace	transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))	transforms.v2.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))	Should use v2 transforms namespace for consistency if migrating to v2.
transforms.Compose() from v1 namespace	transforms.Compose([...])	transforms.v2.Compose([...])	Should use v2 transforms namespace for consistency.

Suboptimal / Outdated Patterns

Issue	Current Code	Modern Alternative	Notes
Device selection only checks CUDA	torch.device('cuda:0' if torch.cuda.is_available() else 'cpu')	torch.device(torch.accelerator.current_accelerator().type if torch.accelerator.is_available() else 'cpu')	Does not account for MPS or XPU. torch.accelerator API available since PyTorch 2.4.
Model save uses .pth extension	torch.save(net.state_dict(), './cifar_net.pth')	Use .pt extension: './cifar_net.pt'	PyTorch convention now prefers .pt over .pth to avoid confusion with Python path config files.
Links to GitHub master branch	https://github.com/pytorch/examples/tree/master/imagenet	https://github.com/pytorch/examples/tree/main/imagenet	PyTorch repos have migrated default branch from master to main.

Files

• beginner_source/blitz/cifar10_tutorial.py

[saurabhkthakur]

/assigntome