Add FDATransform

[guarin]

Add FDATransform based on Disentangling the Effects of Data Augmentation and Format Transform in Self-Supervised Learning of Image Representations.

The different subtransforms have already been added in https://github.com/lightly-ai/lightly/issues/1646 The goal now is to combine them all into a single FDATransform. Note that the image in the paper shows that the amplitude rescale and phase shift transforms are applied in parallel but we can implement them sequentially as the subtransforms handle the split into amplitude/phase components internally.

Tasks

• [ ] Add FDAView1Transform that combines transforms from https://github.com/lightly-ai/lightly/issues/1646
• [ ] Add FDAView2Transform that combines transforms from https://github.com/lightly-ai/lightly/issues/1646
• [ ] Add FDATransform that combines FDAView1Transform and FDAView2Transform

Follow the BYOLTransform code on how to implement the different views and the full FDATransform: https://github.com/lightly-ai/lightly/blob/master/lightly/transforms/byol_transform.py

Each view transform must implement the following steps:

Parameters for view 1 and 2:

[vectorvp]

Hello @guarin, please assign this issue to me.

[guarin]

Thanks for looking into it!

[vectorvp]

@guarin, while implementing this I see there is no probability parameters in AmplitudeRescaleTransform and PhaseShiftTransform. Should we add it or what is the workaround?

[guarin]

Hi, you can use the RandomApply transform from torchvision