Turns out the solution proposed in #27 to adapt the usage of log_abs_det_jacobian for torch>=1.8 was not ideal -- in particular, we only need to use the IndependentTransform when we do not request automatic transforms. The way we set up priors for the tasks, event dimensions are already declared. Wrapping twice will lead to unexpected behavior. This PR fixes this problem and adds new tests to detect related issues in the future.
Turns out the solution proposed in #27 to adapt the usage of
log_abs_det_jacobian
for torch>=1.8 was not ideal -- in particular, we only need to use theIndependentTransform
when we do not request automatic transforms. The way we set up priors for the tasks, event dimensions are already declared. Wrapping twice will lead to unexpected behavior. This PR fixes this problem and adds new tests to detect related issues in the future.