Add Rice distribution #4
Conversation
|
This implementation uses Rician functional forms for all the attributes. However, it is noteworthy that when nu >> sigma, the Rice distribution becomes indistinguishable from a normal distribution with mu=nu,sigma=sigma. There are further details in the wikipedia artical. My own experiments suggest that switching to a normal PDF when nu/sigma > 40 improves numerical stability. Likely this cutoff could be much lower without impacting accuracy. Further study could be beneficial here if we run into numerical issues with this distribution implementation. |
|
Note there were some breaking changes to the setup-python action for MacOS. This PR will also change the test workflow to support python 3.8/3.9 on macos-13 rather than macos-latest. |
| log_z, log_nu, log_sigma = torch.log(z), torch.log(nu), torch.log(sigma) | ||
| log_a = log_nu - log_sigma | ||
| log_b = log_z - log_sigma | ||
| ab = torch.exp(log_a + log_b) # <-- argument of bessel functions |
There was a problem hiding this comment.
again, any reason why not just:
ab = nu * z / (sigma * sigma)
|
going ahead and merging this pr to restore CI and add the Rice distribution. as discussed offline with @minhuanli, the tests in this pr do not assess the validity / accuracy of the implicit reparameterization gradients. we should research a method of testing those cheaply and programmatically across the API. i added #8 to remind us of this shortcoming. |
This PR adds a learnable Rice distribution which is useful in modeling structure factors in X-ray crystallography.