Add MDN with support for parameter transformations
alvorithm opened this issue · comments
Building on Marcel's work, add an MDN (Mixture Density Network) implementation that can play well with e.g. SNPE-A proposal posterior corrections.
This is the wishlist
- basic MDN that works and is useable from sbi (mixture of full-covariance Gaussians)
- MDN becomes highly compositional - only last layer implemented (MDN layer), the rest is built e.g. using
Sequential
and possibly heuristics given data dimensions, etc. - MDN layer returns a
MixtureSameFamily
density (becoming part of PyTorch, PR) - (long term) MDN layer can self-configure given the particulars of the desired
MixtureSameFamily
- Building the whole MDN does not require specification of redundant information, much like Keras s sequential (or look at thinc for a more functional take).