Allow training on customized dataset

[duyao-art]

I revised the original train.py. I am a new guy in this field, therefore any suggestions are welcomed!

Check README_customized_data.md to see the training process.

The updated train_SSL.py allows training using customized dataset. The original MixMatch Loss Function was also revised to consider the potential class imbalance issues during the training process.

The updated train_TL.py allows easier perfomance evaluation of SSL and TL training paradigm.