You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Computing the CDF of a normal distribution gives the wrong result if the location is specified as an integer array, and CDF is called with the smallest legal int32 value.
MWE: The following should all return 0.0, as the CDF $F(x)$ of a Gaussian should approach 0 for $x \to -\infty$
A workaround for me is to explicitly cast the location vector to float.
This is a highly unlikely edge-case, but it did occur for me during a real-world problem (computing mutual information).
System info (python version, jaxlib version, accelerator, etc.)
Description
Computing the CDF of a normal distribution gives the wrong result if the location is specified as an integer array, and CDF is called with the smallest legal int32 value.
MWE: The following should all return$F(x)$ of a Gaussian should approach 0 for $x \to -\infty$
0.0
, as the CDFBut instead it yields:
Note that
log_cdf
seems to be OK.A workaround for me is to explicitly cast the location vector to float.
This is a highly unlikely edge-case, but it did occur for me during a real-world problem (computing mutual information).
System info (python version, jaxlib version, accelerator, etc.)
The text was updated successfully, but these errors were encountered: