Amrit Krishnan

One should always use a validation set if you are tuning stuff, and also to notice overfitting. So, you can take a small set out of the training say 10k...

don't think so, the API has changed quite a bit. For example, the measurements data structure has changed, so the controller won't work.

Do you mean the ground truth disparity maps? The KITTI benchmark provides it http://www.cvlibs.net/datasets/kitti/eval_stereo.php.

> Instead of returning sometimes returning `NoneType`, would it be better to check if it's `None` before `isinstance`? yeah that sounds better. i wonder if the metric_dict is the only...

Is there some direction regarding this from the HF team @lewtun ? Can conversion from polars to HF dataset be implemented with limited/zero copy? So, something like ``Dataset.from_polars()`` and ``Dataset.to_polars()``...