remove script annotations / eager compilation

[eellison]

After pytorch/pytorch#38050, classes compile recursively so there is no longer any need to annotation them. And we can use _script_if_tracing to not eagerly compile the script functions but still compile them when we trace.

I'm hoping this resolves #2132

Edit looks like #38050 hasnt made its way to nightly yet

[fmassa]

LGTM once tests pass, thanks a lot @eellison !

[fmassa]

@eellison test failures seems related, could you have a look?

[fmassa]

Lint failures are unrelated and have been fixed in master. Thanks a lot Elias!

[IRDonch]

It's unfortunate that this wasn't included in 0.6.1. Since future releases will apparently only support Python 3.6+, if you just run pip install torch torchvision on Python 3.5 you'll still end up with a broken configuration.

I don't suppose there's a 0.6.2 planned?

[fmassa]

@IRDonch There isn't, but we are going to be releasing 0.7.0 in the next few weeks (but no Python 3.5 support)

[fmassa]

Also, @IRDonch note that this is only broken for torchhub models, the rest of torchvision should not be broken for the 0.6.1 release on Python 3.5

[IRDonch]

the rest of torchvision should not be broken for the 0.6.1 release on Python 3.5

Just importing torchvision fails for me (with the same error as in #2132), even with torch==1.5.1 torchvision==0.6.1.

[fmassa]

@IRDonch are you using virtualenv? It had some bad interactions with @torch.jit.script annotations, but it has since then been fixed in master and the fix will be present in 0.7.