@pytorchbot merge this please

@ezyang merged this pull request in f8cab38.

Is this error fixed? I am getting this error:

---------------------------------------------------------------------------
RuntimeError                              Traceback (most recent call last)
<ipython-input-8-7babe3f1c55d> in <module>
      1 mu2 = mean1
      2 std2 = Cov1
----> 3 q = MultivariateNormal(mu2, std2)

/projects/sysbio/projects/rojin/anaconda3/lib/python3.6/site-packages/torch/distributions/multivariate_normal.py in __init__(self, loc, covariance_matrix, precision_matrix, scale_tril, validate_args)
    140             if precision_matrix is not None:
    141                 self.covariance_matrix = torch.inverse(precision_matrix).expand_as(loc_)
--> 142             self._unbroadcasted_scale_tril = torch.cholesky(self.covariance_matrix)
    143 
    144     def expand(self, batch_shape, _instance=None):

RuntimeError: cholesky_cpu: For batch 0: U(13,13) is zero, singular U.

@rojinsafavi This PR enhances the stability for precision_matrix. It look like your std2 is not positive definite. Maybe you can solve it by adding a small term to the diagonal part of std2.

So I'm actually using the same thing that I use in "R" kldiv function, and it works fine over there, is something different here?

@rojinsafavi This PR enhances the stability for precision_matrix. It look like your std2 is not positive definite. Maybe you can solve it by adding a small term to the diagonal part of std2.

This is the diagonal, no negative number:

array([0.5891058 , 0.7394688 , 0.84647965, 0.7394722 , 0.77156126, 0.5949623 , 0.7069786 , 1.4996048 , 0.49080923, 0.90659595, 0.29516238, 0.83300734, 0.81732637, 0.71747035, 0.6509524 , 0.797131 , 1.103374 , 0.5067505 , 1.1981595 , 1.0099905 , 0.9176539 , 0.70608026, 0.42208812, 1.0714566 , 0.7964946 ], dtype=float32)

I think positive diagonal is not equivalent to positive definite. You can check if std2 is positive definite in single precision (float32) by using torch.eig(std2) and checking if the eigenvalues are positive.

I think positive diagonal is not equivalent to positive definite. You can check if std2 is positive definite in single precision (float32) by using torch.eig(std2) and checking if the eigenvalues are positive.

Thanks for the help, I was not using the right covariance matrix. The result is a bit different from what I get in R though. In R the KL is 1.955039, and in torch it is tensor(1.9948). Just to make sure that I'm doing everything as I should, here are the steps:

mu1 = mean0
std1 = Cov0
p = torch.distributions.MultivariateNormal(mu1, std1)
​
mu2 = mean1
std2 = Cov1
q = MultivariateNormal(mu2, std2)
​
torch.distributions.kl_divergence(p,q).mean()
tensor(1.9948)

Cov0.txt
Cov1.txt
mean0.txt
mean1.txt

@rojinsafavi I think the way you use kl_divergence is correct.