datasets: Fallback to our own mirrors for mnist

[seemethere]

We are experiencing 403s when trying to download from the main mnist
site so lets fallback to our own mirror on failure.

Re-do of #1940

We have now gotten permission to mirror the MNIST dataset

Signed-off-by: Eli Uriegas eliuriegas@fb.com

[pmeier]

One minor nitpick below. Still, could you give an example where you are experiencing an 403 error? Our download test pass for MNIST (EMNIST fails but with an unrelated issue) and running

from torchvision import datasets
datasets.MNIST(".", download=True)

also works for me locally.

[pmeier]

These seem like debug statements, no?

Suggested change

	finally:
	print()

[fmassa]

I think this might be here to avoid the download bar stopping at the middle of the screen for some reason. I'd be happy to remove it in a follow-up commit if this is unnecessary, but I'll be moving forward with merging this as this is blocking a few other things

[Zethson]

@pmeier https://github.com/pytorch/vision/runs/2084843908?check_suite_focus=true

[pmeier]

@Zethson The requests fail with an 503 instead of 403 error. Meaning the server maybe just undergoes maintenance, but the access is not generally forbidden. @seemethere do you have insights?

[fmassa]

Thanks @seemethere !

Let's get this merged now to unblock, and maybe revisit @pmeier comments in a follow-up PR

[fmassa]

I think this might be here to avoid the download bar stopping at the middle of the screen for some reason. I'd be happy to remove it in a follow-up commit if this is unnecessary, but I'll be moving forward with merging this as this is blocking a few other things

[pmeier]

@seemethere @fmassa

This regressed the download of (K|Fashion)MNIST variants. They rely on the download functionality of the original MNIST. Given that hosting MNIST ourselves is not a temporary solution, couldn't we simply remove the original download links? Otherwise we need to refactor this module.

[fmassa]

@pmeier let's fix the MNIST variants.

We got permission to mirror if the download to the original server fails, so we shouldn't just remove the original links