DCGAN-parametrized ptychography example

Our ultimate goal is to detect patterns while reconstructing the volume, to maximally correlate information spatially and drastically improve our data efficiency. The key to correlation is by using a DCGAN.

We technically can do this without the Kirkland parametrization by outputting the fine-scale potential directly from the DCGAN, and that may be nice to have anyway.

Plan

Once #17 (closed) is closed, we will be ready to create the species and position variables using a neural network. We should then follow [1] or similar to generate a very simple upscaling CNN and introduce a new lower-resolution parameter that we will estimate along with the CNN weights.

We should gracefully handle cases where --kirkland is not provided but --dcgan is.

[1] https://pytorch.org/tutorials/beginner/dcgan_faces_tutorial.html