offer to best model for preTrainDArch

[alex7tula]

Function preTrainDArch return last model.
May be is better to return best model + add stopAfterEpochs?

[saviola777]

I may add options for that in the future, but I don't think it has huge importance or impact, because the pre-training is unsupervised and mainly used to initialize the network weights in a way that allows fine-tuning to quickly converge, the actual error during pre-training is not so important.

[alex7tula]

If to start from the best starting position, we can faster and better trained in the final stage

[alex7tula]

Otherwise it makes no sense to do pretraining more than 1 epoch.
10th epoch as accidental as the first.