BUG: data access issue for the DL tutorial

[bsipocz]

CI cron run into data access issues with the DL notebook. A series of restarts fixed most of the jobs, but not all, and anyway it should not be a problem when there is only a handful runners trying to grab the same data (e.g. the problem will be way more present when it's a full room of workshop attendees).

So I open this issue as a reminder, and if this is not a one-off problem, to do something about it.

[rossbar]

Yeah looks like 429 errors, i.e. server request limits. Right now the data is hosted on github (in a personal repo of mine 😱 ) which was our "temporary" solution to the 503 errors when we were pinging the server on which the data was originally hosted.

I would be very surprised if the MNIST data weren't already hosted somewhere publicly and more sustainably, so we should investigate + switch to that!

[melissawm]

Total overcomplication but if we can't find another alternative, torchvision packages a version of mnist: https://docs.pytorch.org/vision/stable/generated/torchvision.datasets.MNIST.html

[bsipocz]

If I see correctly they are still just grabbing from these two addresses:

    mirrors = [
        "https://ossci-datasets.s3.amazonaws.com/mnist/",
        "http://yann.lecun.com/exdb/mnist/",
    ]

[bsipocz]

(but at least the s3 one should be resilient for multiple downloads, so I would be +1 for swapping over the uris. However, I would not like adding pytorch as a dependency just for making use of this function)