Gamma distribution?

[andywong36]

Is there a plan to add support for a gamma distribution GLM in the future?

The current list of supported GLM distributions are:

['bernoulli', 'binomial', 'categorical', 'multinomial', 'poisson', 'negative_binomial', 'exponential', 'weibull', 'gaussian', 'ceiling_weibull']

There are also a large number of transforms defined in foundry/glm/glm.py:66-120. Have these been validated? Can we assume that we're actually doing MLE of the parameters in the GLM (assuming penalty=0.), or are we optimizing for some other objective?

[jwdink]

Most of the distributions in torch.distributions are supported by just creating your own "Family" instance. Adding support for other aliases (eg being able to pass a string "gamma" to Glm) is also super easy -- feel free to create a branch/PR for this.

I'm not sure I understand your second question, but yes I believe the optimized params should be the MLEs if penalty=0

[jwdink]

I thought about this in a059f63 but there's actually two complications:

• We should try to make these aliases use "location" parameterizations, meaning the default (only predicting the first-parameter of the distribution with covariates) allows predicting the mean of the distribution. This led to the current implementation of NegativeBinomial vs. the default. I think this should be pretty easy with gamma, but haven't done it yet.
• Gamma doesn't implement a CDF, but it could!: Gamma distribution does not have a CDF implementation, which is very useful in Pyro pytorch/pytorch#41637 (comment)