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
test_adjoint.py::test_sequential_sum_product_adjoint: xfail_param(MarkovProduct, reason="mysteriously doubles adjoint values?") MarkovProduct is evaluated again at actual.reduce(sum_op) line:
withAdjointTape() asactual_tape:
actual=impl(sum_op, prod_op, trans, time, {"prev": "curr"})
actual=actual.reduce(sum_op)
It might be due to " .reduce binds variables, which triggers alpha-renaming and reinterprets actual?" from slack discussion with @eb8680
MarkovProduct is not expanded by lazy (or any other Funsor not defined in adjoint_ops). A temporary solution (Adjoint markovproduct #492 ) might be to define an adjoint for MarkovProduct. Still work in progress.
This is to track issues related to using adjoint.
test_adjoint.py::test_sequential_sum_product_adjoint
:xfail_param(MarkovProduct, reason="mysteriously doubles adjoint values?")
MarkovProduct
is evaluated again atactual.reduce(sum_op)
line:It might be due to "
.reduce
binds variables, which triggers alpha-renaming and reinterpretsactual
?" from slack discussion with @eb8680MarkovProduct
is not expanded bylazy
(or any other Funsor not defined inadjoint_ops
). A temporary solution (Adjoint markovproduct #492 ) might be to define an adjoint forMarkovProduct
. Still work in progress.TraceEnum_ELBO
usesAdjointTape
to calculate expectations. Undereager
evaluationnan
gradients appear in more than half of the tests as caught by Catch nan values in TraceEnum_ELBO grads pyro#2782:The text was updated successfully, but these errors were encountered: