Closed matt-graham closed 1 year ago
Looking at the actual numerical differences this could just be due to floating point errors arising from computation of summations in different orders. The relative difference (113431310.40270501 - 113431310.40270479) / (113431310.40270479)
is of order 1e-15
which is close to machine epsilon for double precision floats. We may just need to slightly relax the assertion to use something like numpy.allclose
.
One test failure still when running with Pandas 2.0 / Python 3.11
Originally posted by @matt-graham in https://github.com/UCL/TLOmodel/issues/1114#issuecomment-1728951337