Freeze model performance

[ChrisZhangcx]

Freeze model performance in multiple training process. The simplest way is to freeze one-hot encoding for each class every time we re-run the train.py. By applying this, we can freeze the model performance to 0.8250 with loss 0.7380.