Provide path_derivative argument to all continuous distributions about zhusuan HOT 5 CLOSED

Thanks @botev . This is good since it just keeps existing code behaviors unchanged. I've thought about this sticking-landing estimator but was focusing on how to do it in the variational objective which requires change of the api. Do you have interest in submitting a PR? We should appreciate it.

from zhusuan.

Sure, I'll be happy to make a PR.
Not use what kind of tests would be needed though?

from zhusuan.

A simple test that checks the gradient actually doesn't go through the log_p term in that way would be enough (e.g., log_p has no gradient w.r.t. parameters if givens are not samples). And if I'm right you only need to implement this for distributions that has a is_reparameterized argument (Normal, Concrete, etc.).

from zhusuan.

Ok and lastly do you want to have use_path_derivative as an argument to log_prob and _log_prob, to be as an attribute in the general Distribution class (like is_reparametrized) or attribute in the concrete classes. I think it should be an attribute to log_prob and _log_prob in the abstract Distribution class.

Also, could you point me out to how to compare the numeric results in your test suites? E.g. to verify it I will explicitly calculate the gradients and want to assert that they are all close to the correct ones.

from zhusuan.

I feel it's good to have it as an attribute to the Distribution class but you may need to actually do the stop_gradient in each concrete class because their parameters vary, just like how the is_reparameterized argument is dealt with.

from zhusuan.