170592 is completely wrong (negative all the way with autoregression, huge error bar without autoregression). 176462 is right but around the end of the shot, the non autoregressive model flips. Might be a sign error in some of the data.
Would suggest making a plot of rotation vs tinj over the whole database (should be both positive or both negative generally). There are cases where torque flips and therefore rotation flips 175970
From Hiro:
170592 is completely wrong (negative all the way with autoregression, huge error bar without autoregression). 176462 is right but around the end of the shot, the non autoregressive model flips. Might be a sign error in some of the data.
Model without autoregression during training:
Model with autoregression during training: