Retraining a new model so that we have a matching .pth file.
Also, discarding more data on init to get better randomization.