What decoder transforms should we work on?

[scotts]

We currently support three decoder transforms:

1. Resize
2. CenterCrop
3. RandomCrop

The purpose of this task is to start the conversation of what additional decoder transforms folks would find useful. Please provide feedback! Some potential work to do:

• More options for Resize. Currently, interpolation is always bilinear and antialiasing is always on.
• The padding options for RandomCrop.
• RandomHorizontalFlip, maps to FFmpeg's hflip.
• RandomVerticalFlip, maps to FFmpeg's vflip.
• Pad maps to FFmpeg's pad.
• RandomRotation maps to FFmpeg's rotate.
• Grayscale maps to FFmpeg's colorchannelmixer.

The above are, I think, the "easy" transforms where I'm confident there is a straight-forward mapping. There's also a possibility for us to implement support for an FFmpeg transform that does not have a direct analog in TorchVision, but we'll want to make sure it won't be confused with a TorchVision transform.

[scotts]

A few other things to consider:

• The FFmpeg filter zscale is an alternate to scale. We may need to use it in Resize if we change our output formats to something other than RGB24.
• Audio transforms! We haven't talked about those, but we should be able to implement a similar relationship with TorchAudio's transforms.