[C++ API] Support for CIFAR10 and CIFAR100 Datasets

[prabhuomkar]

🚀 Feature

C++ API should support Famous Open Datasets CIFAR10 and CIFAR100.
Dataset Reference: https://www.cs.toronto.edu/~kriz/cifar.html

Motivation

While creating PyTorch C++ tutorials https://github.com/prabhuomkar/pytorch-cpp, I came to know that PyTorch has support for these datasets via vision. As of now, C++ API for Datasets only has MNIST.
Support for standard datasets like CIFAR, COCO, ImageNet, etc out of the box will help developers play with C++ API easily.

Pitch

• Add cifar.h header with TORCH_API CIFAR class.
• Add cifar.cpp with both CIFAR10 and CIFAR100 support similar to vision for reading images, targets and getting train/test data.

Alternatives

N/A

Additional context

Reference C++ Implementation: cifar10.h and cifar10.cpp
Reference Python Implementation: cifar.py

cc @yf225 @fmassa

[prabhuomkar]

cc @yf225 @soumith @apaszke

I would like to contribute to this Issue.

[yf225]

@prabhuomkar Thanks a lot for the suggestions! My personal take is that we should put the C++ vision datasets in https://github.com/pytorch/vision, same as the Python ones. @fmassa Curious what are your thoughts on this?

[prabhuomkar]

@fmassa would like to know your thoughts on this before closing this.

[soumyadipghosh]

@prabhuomkar Did you finally implement support for CIFAR-10/ImageNet in C++ in torchvision?

[prabhuomkar]

@soumyadipghosh
Not yet, but I have a reference CIFAR implementation:
https://github.com/prabhuomkar/pytorch-cpp/blob/31d325e34c0d3d3578dd5e811a2b459a78dcc32f/tutorials/popular/blitz/training_a_classifier/include/cifar10.h
https://github.com/prabhuomkar/pytorch-cpp/blob/31d325e34c0d3d3578dd5e811a2b459a78dcc32f/tutorials/popular/blitz/training_a_classifier/src/cifar10.cpp

Will be creating PR very soon!

[soumyadipghosh]

@prabhuomkar Thats great! Are you aware of anything similar for ImageNet?

[prabhuomkar]

Not yet, will try to add one by this weekend as an example

[soumyadipghosh]

That would be nice!

[prabhuomkar]

@soumyadipghosh use this prabhuomkar/pytorch-cpp#73 to track?

[soumyadipghosh]

Ok