transforms: randomly grayscaling an image

[sourabhd]

Added transform for data augmetation technique where an image is grayscaled with probability p. The grayscaled version of image can be three channel (r == g == b) or single channel.

For detailed discussion of this feature please refer to issue #299

[alykhantejani]

LGTM. Can you add some simple tests to check that to_grayscale returns the expected number of channels and data?

[alykhantejani]

LGTM. Can you add some simple tests to check that to_grayscale returns the expected number of channels and data?

[sourabhd]

@alykhantejani

1. I have added test cases
2. Subtle change to the interface:
   F.to_grayscale() remains same. We have two transforms:

• Grayscale(num_output_channels=1) : converts RGB/1-channel grayscale to 1 channel or 3 channel grayscale as desired
• RandomGrayscale(p=0.1) : randomly grayscales image, but retains the number of channels. [Does not make sense to have have few 3 channel and few 1 channel images in a batch]
  Please review the changes again

[alykhantejani]

@sourabhd Thanks for making the changes. There seems to be some merge conflicts with master that need to be resolved before we can merge

[sourabhd]

@alykhantejani RandomRotation and RandomGrayscale seem to be added at the same time in same files. Not a real conflict - just need to move lines of code and it should be fine. Write access is needed for merge, which I do not have.

[alykhantejani]

@sourabhd yup, I can do the merge once the conflict is resolved. Can you do this locally and push to origin? i.e.:

git remote add upstream [email protected]:pytorch/vision.git
git pull upstream master
<this will cause a merge conflict>
<fix conflict>
git commit -am "merge upstream"
git push origin master