Overfit for validation set and test set?

[Gears_Gears]

Hi!

May I get some advice for the following situation please?:
I’m doing cross-validation. In each cross-validation round, I return the lowest training loss and the lowest validation loss of that epoch. Then, I choose the model for testing by lowest validation loss.

The problem is: while my test loss (0.05) is close to the validation loss (0.04), in each validation round, my validation error is always 1.5 to 2 times that of the training error.

My question is:
1 Is this overfitting? It looks like my model overfitting for the cross-validation round (because training loss much smaller than testing loss), but not-very-overfitting for the testing round (because the validation loss is close to testing loss))? This is confusing because the data used for testing and cross-validation are the same. Perhaps the model just generalizes poorly in general?
2 What should I do? My model is for regression, with input being a long 1d vector and output being a single number. (I have early stopping)
I used dropout but that made the model converge at high loss
I tried to ‘strengthen’ the regularization by reducing early stopper patience from 5 to 3, and got higher loss again

[Gears_Gears]

My training loss is blue, and my testing loss is orange.
In each validation round, the loss look like:

It looks my model converged, but by the time I stop it, the model has overfit (because train error drop below the test error). It looks like the only thing I could make the model stop earlier is to just do ‘more’ early stopping

But ‘more’ early stopping (I am counting the number of times the model doesn’t give lower validation loss) means lower patience, and my patience is already 5 (as described above, going down to patience=3 only makes me stop at higher loss, and I don’t think patience=1 is a good idea, since the model’s loss might go up and down)