hi 一起讨论一下输出loss逐渐变成nan值情况的改进策略Let's discuss the improvement strategy for the situation where output loss gradually changes to nan value together

[love112358]

大家好！这是我使用自己采集音频数据训练的结果，刚开始的loss值并不是nan，最后，当训练到120个epoch时，loss值变成nan值
Hello everyone! This is the result of training using audio data that I collected myself. Initially, the loss value was not NaN, but eventually, when training reached 120 epochs, the loss value became NaN.
epoch=122

epoch=123

epoch=149

这是tensorboard的显示结果
This is the display result from TensorBoard.

希望大家给我一下赶紧的建议
I hope everyone can give me some urgent advice.

[yl4579]

see #60