A set of implementations of different image transform codes (sparse coding, ICA, PCA, etc.) to facilitate comparison between them. The implementation is in the PyTorch GPGPU framework and should be relatively performant in terms of wall-clock time for training and inference. The modular organization is to stress the interchangibility of different techniques for code inference (analysis) and reconstruction (synthesis).
This also includes some basic image processing utilities that may be useful to someone working with images in Python, particularly when fitting transform codes to image datasets.
Of particular interest is code for doing convolutional sparse coding and for enforcing "subspace" or "topographic" constraints on sparse codes. The convolutional variants of sparse coding are implemented in the convolutional/ sudirectories of analysis_transforms and dict_update_rules. The subspace analysis transform is implemented in analysis_transforms/fully_connected/subspace_ista_fista.py. The topographic sparse coding implementation lives on the topographic-sparse-coding branch, I haven't yet merged it into master.
Spencer Kent