vocalpy
ENH: Add RandomCrop transform
#555 proposes to deprecate WindowDataset.
One good thing about the WindowDataset abstraction is that kind of gives us data augmentation and thus translational invariance "for free", because a network can see literally every possible window in the dataset.
We don't want to lose that.
To achieve something similar, we should have a RandomCrop transform.
as described in #169:
with a "random crop" type transform where we just take a single window from each item in the dataset, where the x for each item is now the whole spectrogram instead of the current abstraction where x is a window
Importantly we will need the ability to make RandomCrap window size aware. We want it to crop but in such a way that an individual sample in the dataset can be converted to a set of consecutive non-overlapping windows. Part of the way to achieve this is with cropping. Need to think more about that.
Other desiderata:
- [ ] ideally this crop would be "smart" -- for the case of SED one could configure how it crops, e.g. only within silent intervals of a specified duration?
- [ ] also nice-to-have would be logging what window was used, e.g. through
tensorboard - [ ] and then being able to re-run an experiment with the exact same windows ... this is "extreme reproducibility" but would be nice to rule out any effect of window choice
