Fix dataset url

[lanpa]

There is "HTTP Error 500" for http. change url to https fixes this.
(link in the official cifar page also says https://)

[soumith]

The http works fine for me, i just checked...

[apaszke]

It might work, but it's better to use https if we can

[soumith]

no, https is giving subtle problems based on user machines and python version. lets stick to http please (we verify hash anyways)

[apaszke]

It's md5, so it's not really secure.

[soumith]

@lanpa is it still happening for you? or is the 500-internal-error an intermittent message you get sometimes?
@apaszke let's switch to a more secure hash than md5 then.

[lanpa]

@soumith You are right, that error happens occasionally. How about

try 
    download_url(http_url, ...) 
except 
    download_url(https_url, ...)

to handle them all?

[soumith]

yes i think that's a good idea.
Actually the reverse:

try:
https_url
except:
http_url

[soumith]

thanks @lanpa !

[alykhantejani]

@soumith @apaszke should we update this download function in the same way that we did for pytorch's load_url, I think that'd be cleaner and more consistent?

[lanpa]

@alykhantejani What if we also move model_zoo.py from torch to torchvision? Because it does not referenced by functions in torch, the only invokers lie in torchvision/model/*.py.

[fmassa]

@lanpa originally model_zoo.py was in torchvision, but we decided to move it to pytorch because it can be used elsewhere (like torchtext for example). I think it's better to leave it there.