About loss issue

[laoniandisko]

I trained your source code, and after training for 10epochs, the loss basically does not change, it is still very large, and the test results are very bad. I would like to know if your pretrained model has any structure to change the code

[vivektyagiibm]

I agree with @laoniandisko observation and I also ran into the same problem. The loss gets stuck and the recognition results are very poor. There seems to be a bug in training code.