Physical Datasets, Naive and Flow, v03 Translation

This is with 2 different translation datasets, and the flow vs naive methods.


Using data augmentation gaussian_0.00001.
Remind, eval MSEs are not directly comparable due to tool flow vs direct vector but we might want to change this.
We do NOT do any scaling for the flow-based methods because I figured (a) this was over-complicating things, (b) if we do scale, and we predict the scaled flow vectors, the induced rotation from that will not be the same as the induced rotation if we divide the scale vectors, so that's another problem (though in this case, not really an issue since it's translation only datasets).

Results

Just the Flow Methods