MaoYuxin

OK, I will try again after modifying it, thank you. In addition, I checked the log and found that NAN will appear at epoch=5 during my training process.

After setting `--num_classes 10` and disabling the modulate kernel, training is now normal. In the step where NAN occurred before, NAN did not occur. Thank you for your positive reply....

Hello, do you have any plans to open source the training code? I look forward to your answer.