try to improve val loss stability

[jpata]

• instead of pt, predict and run loss on log(pt/elem pt), as this quantity is distributed approximately as a Gaussian (same for E)
• save a few validation events at each epoch for checks to understand why the validation loss is fluctuating a lot