Giters

probtorch / pytorch

Tensors and Dynamic neural networks in Python with strong GPU acceleration

Home Page:http://pytorch.org

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

kl_divergence(Bernoulli(0), Bernoulli(0)) == NAN

fritzo opened this issue · comments

commented

It looks like we could improve numerical stability of kl_divergence().

  • kl_divergence(Bernoulli(0), Bernoulli(0)) == 0 (currently NAN)
  • kl_divergence(Bernoulli(0), Bernoulli(1)) == flaot('inf') (currently NAN)
  • Categorical
  • OneHotCategorical
  • Poisson
commented

@vishwakftw Do you want to look into this?

commented

@fritzo Sorry about the Bernoulli distribution. I have fixed it locally, and currently I am waiting for the build to complete for testing. W.R.T. the other, could you suggest what needs to be done?

commented

No worries 😄 I'm glad we could get the implementations merged early so we can test things in Pyro.

I think it would be good to have tests of edge cases as in Neeraj's test_bernoulli_gradient. Let's add something like TestKL.test_edge_cases() with a few tests like

class TestKL(TestCase):
    def test_edge_cases(self):
        self.assertEqual(kl_divergence(Bernoulli(0), Bernoulli(0)), 0)
        self.assertEqual(kl_divergence(Bernoulli(0), Bernoulli(1)), float('inf'), allow_infinity=True)
        self.assertEqual(kl_divergence(
            Categorical(variable([0, 1])),
            Categorical(variable([0.5, 0.5]))), float('inf'), allow_infinity=True)
        ...
commented

Fixed. Closing in favour of pytorch#4961.