Closed
Description
Following my (slow and steady) progress here, check_logcdf
in test_distributions.py
seems to be having issues only in float32
. Below is a summary of the issue in code.
import aesara
import pymc3 as pm
import scipy.stats.distributions as sp
# scipy reference
print(sp.chi2.logcdf(0.01, 100))
# output: -413.39853723533787
# float64
with pm.Model() as model:
print(pm.ChiSquared.logcdf(0.01, 100).eval())
# output: -413.39853723533787
# float32
aesara.config.floatX = "float32"
with pm.Model() as model:
print(pm.ChiSquared.logcdf(0.01, 100).eval())
# output: -inf
notebook available here
Happy to hear any pointers or suggestions!
Metadata
Metadata
Assignees
Labels
No labels