Training and Validation Errors #329
-
|
I have noticed that my training set errors are consistently larger than my validation set errors when I train a model. Is this due to the model not re-scaling into real units, similar to the answer to another question about the difference between model.train() and model.eval()? |
Beta Was this translation helpful? Give feedback.
Replies: 1 comment
-
|
Hi @tjwolter , I'm struggling right now to find the relevant GitHub discussion, but maybe you'll have better luck than me: this is a consequence of our use, by default, of EMA averaging on the validation model (but not the training model) and is generally expected. All reported errors are in real units, regardless. (Losses I think it depends, I'd have to check again and remind myself.) |
Beta Was this translation helpful? Give feedback.
Hi @tjwolter ,
I'm struggling right now to find the relevant GitHub discussion, but maybe you'll have better luck than me: this is a consequence of our use, by default, of EMA averaging on the validation model (but not the training model) and is generally expected. All reported errors are in real units, regardless. (Losses I think it depends, I'd have to check again and remind myself.)