How does the overtraining detector work?

[litsa-the-dancer]

As of recently I've noticed you guys added an overtraining detector thingy. I'd like to know how it works and what graphs and values it takes into account.

[blaisewf]

You can check this commit for more info https://github.com/IAHispano/Applio/commit/5593d183ed44d24639667ecb385fe14088b71693

In each epoch it stores the loss gen all of that point and from there it makes the comparisons.

[litsa-the-dancer]

So that's how it finds the global minimum? If so that really isnt all that accurate, its better for the model to look at loss/g/total rather than loss/gen/all. Looking at loss/g/total gives u a much better representation of how the generator is doing with the training data whilst also taking into account all the other factors such as mel, kl and fm.

[blaisewf]

[litsa-the-dancer]

I'll try testing it out