Closed sethaxen closed 1 year ago
Uhm for the next time, this should have been a breaking release, no? There could be clients running production depending on the old and broken implementation. They now have to adjust or have no output at all, which means this change is breaking.
depending on the old and broken implementation.
According to semver, bug fixes are non-breaking. The only problem I see here are the now-failing examples in the OP - but it seems they could be fixed by improved definitions of bijector
for product distributions or an improved check that also supports TruncatedBijector
s with infinite bounds.
As noted in https://github.com/TuringLang/Bijectors.jl/issues/220#issuecomment-1409115410,
ordered(d)
does the incorrect thing unlessd
is unconstrained. This PR adds a check that throws an error if this is not the case. Note that examples like this will now error even though they are fine to use withordered
, since they don't return aIdentity
bijector: