Variable referenced before assignment in 3D KDE

[lpsinger]

Variable referenced before assignment in 3D KDE

[lpsinger]

I just got this again with event 833111 from 2016. I get it reliably when I run with --enable-distance-map and set --seed=0. It's very odd how many events fail when I set --seed=0. I wonder if somehow that seed is an especially bad one and produces very repetitive pseudorandom numbers?

Simply changing bic > best_bic to bic >= best_bic would protect against this exception:

if bic >= best_bic:
    best_means = self.means
    best_assign = self.assign
    best_bic = bic

What, if any, undesirable consequences might this have?

[farr]

I think your suggestion for solving this problem would result in the possibility of no successful trials for the clustering step.

Will

On 29 Sep 2015, at 23:13, Leo Singer [email protected] wrote:

I just got this again with event 833111 from 2016. I get it reliably when I run with --enable-distance-map and set --seed=0. It's very odd how many events fail when I set --seed=0. I wonder if somehow that seed is an especially bad one and produces very repetitive pseudorandom numbers?

Simply changing bic > best_bic to bic >= best_bic would protect against this exception:

if bic >= best_bic:
best_means = self.means
best_assign = self.assign
best_bic = bic

What, if any, undesirable consequences might this have?

—
Reply to this email directly or view it on GitHub.

[lpsinger]

OK. The more immediate cause is that the for the 3D KDE, the BIC is occasionally coming up as -inf because the posterior is identically zero at some of the sample points. Why would that happen? Is this just underflow?

[farr]

If those points are in the clusters that we just cut out of the posterior, then they can be far enough away from the other points that their posterior underflows to zero, yes. I'm not sure what to do about this....

Will

On 30 Sep 2015, at 03:32, Leo Singer [email protected] wrote:

OK. The more immediate cause is that the for the 3D KDE, the BIC is occasionally coming up as -inf because the posterior is identically zero at some of the sample points. Why would that happen? Is this just underflow?

—
Reply to this email directly or view it on GitHub.