Having just downloaded the ILSVRC2017 data, I’m very confused by the folder structure, and how our code is meant to work with it.
It seems that the PyTorch Imagenet example assumes everything is in a simple ImageFolder format, but the actual dataset seems to split the training data into half a dozen folders indexed and labeled with their years, such as “ILSVRC2013_train_extra3”
I highly suspect that I simply downloaded the wrong variant of the dataset, as it seems every vision paper I read agrees on what the “Imagenet1k” dataset is.
I’m wondering where that specific variant is, and what preprocessing is needed to get it into the ImageFolder format.