Added eps attribute to FrozenBatchNorm2d

[frgfm]

This PR aims at tackling #2169 by:

• adding a eps attribute to ̀FrozenBatchNorm2d`
• renaming the n positional argument of FrozenBatchNorm2d constructor to num_features
• adding a unittest to ensure default eps is zero for backward compatibility
• adding a unittest to ensure valid computation when epsilon is higher than zero by comparison with BatchNorm2d forward in eval mode.

[frgfm]

@fmassa I'm not sure where the error comes from, it seems like the tests timed out in test_ops at NMSTester. Do you think it could be due to this PR?

[frgfm]

Also, I made two choices here that any reviewer might want to take a look at:

• renaming the constructor argument n to get closer to BatchNorm2d interface (I didn't see many reasons for someone to use kwargs, but it might be worth mentioning it as a breaking changes if merged into the next release)

• rearranging the forward implementation by switching reshape to view (I might have missed a use case where it's needed), and delaying this reshape/view to the last moment to reduce the LoC. I made tests to ensure the computation is the exact same as previously, and I also checked execution time on both CPU & GPU which were the same.

[zhangguanheng66]

Also, I made two choices here that any reviewer might want to take a look at:

• renaming the constructor argument n to get closer to BatchNorm2d interface (I didn't see many reasons for someone to use kwargs, but it might be worth mentioning it as a breaking changes if merged into the next release)
• rearranging the forward implementation by switching reshape to view (I might have missed a use case where it's needed), and delaying this reshape/view to the last moment to reduce the LoC. I made tests to ensure the computation is the exact same as previously, and I also checked execution time on both CPU & GPU which were the same.

For the first option, I'm not sure if it introduces a BC breaking because it just re-defines the name of the argument.

[fmassa]

Thanks for the PR!

I have a couple of comments, could you look into addressing those?

[fmassa]

Test failures seem related

[frgfm]

Thanks, I fixed those. There is a timeout remaining for windows builds and an unrelated (I think) failure on travis!

[fmassa]

Thanks a lot!