It would be nice if some special math functions would be available in PyTorch with CUDA support.

I am working with neural networks with probabilistic components and as such it's necessary for me to compute the (log-)pdf, which often requires the computation of currently unavailable math functions.

A good start would be to support the gamma, log_gamma, beta, log_beta and factorial functions.

I am aware that I can write my own NumPy-based-modules, but that precludes the use of CUDA rendering training of such models basically impossible (and it's also error prone in terms of derivatives).

As an aside, great work here!