The workaround is applying the solution found on re-generated constraints inside derive_projections, i.e. when solving for the 3rd time -- the relevant variables should be already solved during the 2nd pass i.e. inside finish_inference.
The whole setup will probably get refactored when I work on optimizing shape inference (it's doing a lot of work now).
The workaround is applying the solution found on re-generated constraints inside
derive_projections, i.e. when solving for the 3rd time -- the relevant variables should be already solved during the 2nd pass i.e. insidefinish_inference.The whole setup will probably get refactored when I work on optimizing shape inference (it's doing a lot of work now).