Finally bit the bullet and made the intermediate loss translation actually use different builtin types to keep myself honest. This has exposed all sorts of problems with the translation and it's taken a while to fix them. Finally ditched the non-compositional loss for DL2 as well after Katya's urging.
Next up, fixing your requests for the Python side @wenkokke .
Finally bit the bullet and made the intermediate loss translation actually use different builtin types to keep myself honest. This has exposed all sorts of problems with the translation and it's taken a while to fix them. Finally ditched the non-compositional loss for DL2 as well after Katya's urging.
Next up, fixing your requests for the Python side @wenkokke .