Transform synchronization

[aylward]

Reported by @HastingsGreer @brad-t-moore @floryst

The transforms are synchronized by resetting the random seed, which I guess is fine. Documentation might be needed to clarify that (this is hypothetical)

imtrans=transforms.Compose([
    Rescale(),
    AddChannel(),
    AddGaussianNoise(sigma=.5),
    UniformRandomPatch((64, 64, 64)),
    ToTensor()
 ])    

segtrans=transforms.Compose([
    AddChannel(),
    UniformRandomPatch((64, 64, 64)),
    ToTensor()
])  

wouldn’t work because the AddGaussianNoise layer would desync the random seeds. This would be hard to debug as a beginner!

One guard against this is to check that the random seed is still synchronized after applying the transforms. I put a proof of concept at
HastingsGreer@2295721

Another option is to somehow enforce that random transforms that should apply to both images and segmentations come before random transforms that only apply to images in the compose object.

[wyli]

thank you and the team, I also think relying on the random states to sync the transforms is error-prone in a few ways -- setting random.seed but in transforms using functions from np.random is another example. we're working on a preprocessing pipeline design to address this issue properly.