Fig. 6.

Characteristic training curves of the deep quality prediction network. The plot is an aggregate of all the different validation folds. The validation loss often spikes in early training, which then disappear safter the learning rate reduction at 120 epochs