Feature request: ImageNet data loader

[adrhill]

ImageNet is quite large and locked behind terms of access that require an account.

However it would be nice to be able to either

• set a config (or ENV) variable to download ImageNet through MLDatasets
• point MLDatasets to a local copy of ImageNet

and be able to use MLDatasets' interface of

train_x, train_y = ImageNet.traindata()
test_x,  test_y  = ImageNet.testdata()

as well as ImageNet.convert2image(x). Ideally data would be in WHCN format for Flux and Metalhead models.

[CarloLucibello]

As a reference, an example of ImageNet usage https://github.com/avik-pal/Lux.jl/tree/main/examples/ImageNet

[lorenzoh]

For reference, a ManualDataDep may be useful for when a dataset requires the user to perform some manual steps.

[adrhill]

Thanks for the pointers, I will open a draft PR for this!

Since there is not only one version of ImageNet, I propose to mirror PyTorch and have MLDatasets.ImageNet refer to the ImageNet 2012 Classification Dataset (ILSVRC 2012-2017). The ImageNet authors themselves call it "the most highly-used subset of ImageNet".