Loss vs R Squared (print loss too imprecise)

My regression finished training and has zero loss - training AND validation!

I find that surprising ;), but more to the point the R Squared value is non zero and different for training and validation. Since loss was chosen to be quadratic, how are these facts reconciled?

image

UPDATE - ANSWER

The loss to 2 decimals is missing the significant values…

image

Results should show 2-3 significant figures IMO, not two (truncated, not rounded) decimal places - at the end of the final epoch the training loss was 0.006889…

Hi @JulianSMoore,
Great point! In fact, it should probably say 2.13E-3 to make sure that we don’t loose out on digits :slight_smile:
Added it as a task

I’m Scientific Notation literate - I wouldn’t expect everyone to be, and even so, I prefer explicit digits up to about ~8 because I can see differences in magnitude… scientific notation takes a momentary thought. (Decimals also more accessible for certain sections of the intended user-base?)

So I would suggest that the default is decimal places for all x > 1E-7, always showing 8 digits after decimal point, for example.

Long term, number format could be a user choice…

1 Like