Random rotation

[lkubin]

Description

• We implemented functions for image rotation and random rotation in mxnet.image.image.py, a transform for Rotation and one for RandomRotation in mxnet.gluon.data.vision.transforms.py

Checklist

Essentials

• The PR refers to the relevant issues [Feature Request] Random Rotation Image Augmentation #13206 Random Rotation Image Augmentation #11230
• Changes are complete
• All changes have test coverage
• Unit tests are added for both mxnet.image.image.py functions and for mxnet.gluon.data.vision.transforms.py
• Code is well-documented:
• For user-facing API changes, API doc string has been updated.
• To the my best knowledge, examples are either not affected by this change, or have been fixed to be compatible with this change

Changes

• Image rotation function imrotate, based on BilinearSampler function. Tested in tests/python/unittest/test_image.py
• transforms.Rotate, tested in tests/python/unittest/test_gluon_data_vision.py
• Image random rotation function random_rotate. Tested in tests/python/unittest/test_image.py
• transforms.RandomRotation, tested in tests/python/unittest/test_gluon_data_vision.py

Comments

All the implemented functionalities are based on imrotate function. The shape of the output of this function is the same shape of the input, thus it can be used inside a transformer.
It supports 3 behaviors:

• Simple image rotation without zooming
  
• Zoom in image rotation: no padding is shown in the output result
  
• Zoom out image rotation: the whole original image will be contained in the output result
  

[lkubin]

The validation on windows failed during the build... I think it might need to be restarted.

[lkubin]

@szha I opened this pull request 2 moths ago, I'm wondering if it is an useful feature to be added. Otherwise I'm going to close it.

[marcoabreu]

This is a very useful feature, thanks!

@reminisce @eric-haibin-lin @apeforest could you review please?

[samskalicky]

@lkubin Can you please rebase/merge to fix the conflicts?

[zhreshold]

looks good to me! Can you resolve the conflicts?

[lkubin]

We solved the conflicts and we did the changes that you were suggesting. Let us know if we miss something!

[guanxinq]

LGTM. Thanks for your contribution.

[samskalicky]

@mxnet-label-bot add [pr-awaiting-merge]

[zhreshold]

@lkubin Seems like the ci is consistently failing on windows machines, can you rebase to master and I can look into the problem if it's still failing

[lkubin]

@zhreshold I just merged master into this branch. I'm not sure what's happening with the CI on windows, it seems failing during compiling: I don't think the failure is related to this PR.

[lkubin]

@zhreshold I see that tests on windows are still failing. I cannot understand why though.

[lkubin]

@zhreshold it seems that also the CI on ubuntu on GPU is failing... the process got killed with SIGTERM, maybe a timer expired?

[zhreshold]

Restarted the jobs, let's wait for another round

[lkubin]

@zhreshold It seems that now only linux GPU tests are not passing. Can you restart only those tests to see if they will pass the CI?

[samskalicky]

restarted

[lkubin]

Guys, the CI for Unix still failing, this time with RuntimeError: Python 3.5 or later is required during the installation of pip. Is there something I can do in order to make it work?

[zhreshold]

I think someting is broken in CI which is related to recent deprecation of python2.7

[zhreshold]

@lkubin I think the ci python3.5 error is fixed: #17457

Can you do rebase to master one more time?

[lkubin]

@zhreshold I think that now the build for R is failing...

[samskalicky]

@zhreshold I think that now the build for R is failing...

@lkubin we had to restart the CI last night. Ive restarted the last two CI runs. Hopefully they'll pass now!

[zhreshold]

All green, merged! Thanks guys 😸