arXiv preprint and reference: Relative stability toward diffeomorphisms in deep nets indicates performance (NeurIPS 2021)
Apply a maximum-entropy diffeomorphism to an image img ~ [ch, n, n], in PyTorch.
Usage:
img_diffeo = deform(img, T, cut)
where T is the temperature at which the corresponding displacement fields are drawn and cut is the high-frequency cutoff.
An animated example can be found here.
The range of temperatures corresponding to natural diffeomorphisms - for given n, cut - can be computed by
Tmin, Tmax = temperature_range(n, cut)
To have an idea of how much distortion one gets for given (T, cut), the typical displacement at the center of the image can be computed by
delta = typical_displacement(T, cut)
Samples of max-entropy diffeomorphisms in the (T, c) phase space for an ImageNet sample.
The region [Tmin, Tmax] is colored in green.
