Another reimplementation

[ildoonet]

I got BIG interest on your work so I decided to reimplement with some fixes.

Issue #1 #3 #4 are dealt with this new reimplementation and I still work on improving for better usability and performance.

Currently, after I fixed those issues, I got clearly-improved results on CIFAR100.

	Top-1 Error(@300epoch)	Top-1 Error(Best)
Paper's Reported Result	N/A	13.81
Our Re-implementation	13.68	13.15

I will update as other experiments are updated my implementation progressed.

Repo : https://github.com/ildoonet/cutmix

[hellbell]

@ildoonet
Good.
Please let me know the ImageNet results.
You can send me PR for the implementation.
Thanks!

[ildoonet]

Here is my imagenet experiment result.

	Top-1 Error(@300epoch)	Top-1 Error(Best)
Paper's Reported Result	N/A	21.4
Author's Code(Our Re-run)	21.768	21.586
Our Re-implementation	21.524	21.340

I use your code without modification to replication the paper's result, but I got slightly worse results(21.586). It could be from different environment or other things.

But to me, it is clear that dealing with issues #1 #3 #4 is good for the final performance.

[hellbell]

@ildoonet
Great! It is very happy to see another SOTA result of CutMix from you.
I will verify the modification of #3 and #4 and apply to our code.
Thanks again :)

[hellbell]

Closing the issue.