Is the training stable enough to always converge to the numbers in README.md

[haichaoyu]

Hi, Thanks for your efforts. I used your train.sh and only changed stage to 0. The eval sad turns out to be really high (almost never goes below 80) and fluctuate a lot (epoch 10: sad 86, epoch 11: sad 92). What about your training process, is it stable? Thanks!

[huochaitiantang]

Because there is no stage-0 model trained after 2019.06.17, I will do a new experiment and update README.md later.

[LuchenCatherine]

I encountered such problem before for I didn't use the pretrained model correctly. FYI.