Inverted predictions don't flow cleanly into SaveImage / can't switch collate fn in Invertd
Inverted predictions don't flow cleanly into SaveImage / can't switch collate fn in Invertd: a task in MiMo-V2.6-RL-harbor-code: MiMo-V2.6-RL Code (Harbor) (Harbor dataset). The pre-transforms apply spatial transforms with different parameters per sample (random crops / spacings), so after…
The task
The pre-transforms apply spatial transforms with different parameters per sample (random crops / spacings), so after inversion each sample has a different shape. With the default `collate_fn` in `Invertd` (no collation), `pred_inverted` ends up as a Python list of per-sample tensors without a batch dim — which makes…
Part of FineEnvs/MiMo-V2.6-RL-harbor-code.