Description
-
Is there an official source for the ILSVRC-2012 dataset, which is used for the training process of the models contained in
torchvision
? The documentation points to the official ImageNet homepage, but this seems to be unmaintained. I requested access via the form as well as standard email several times in the last three months, but never got answer. I was using an institute email address, which is not freely available as requested by ImageNet team. -
If the answer to 1. is "no", should we be training the models on ILSVRC-2012 dataset, if the results cannot be reproduced by someone new to the field? Another option could be the ILSVRC-2017 dataset (which AFAIK comprises the ILSVRC-2012 dataset) hosted by the same team on Kaggle. According to the given timeline this competition is open for another 11 years, which guarantees access to the dataset for at least the same time.
-
Is there a reason, why the dataset that the pretrained models are based on, is not part of the
torchvision
package? The documentation hints that this could be implemented asImageFolder
dataset, but I fail to see a reason, why it is not properly predefined. If there is no such reason, I can create a PR to implement this.