How to train better？

Question

How to train better？

1zxLi opened this issue 4 years ago · comments

I did not modify any parameters. After 400K epochs, the loss oscillated around 5.3 and could not continue to converge.
Is there a more suitable parameter setting scheme to continue to optimize the loss? （ljspeech data）

xuexidi · Answer 1 · Fri Sep 25 2020 23:15:26 GMT+0800 (China Standard Time)

I met the same problem...

harryfyodor · Answer 2 · Wed Dec 30 2020 18:38:33 GMT+0800 (China Standard Time)

I met the same problem too, I do not modify anything, but the loss is around 5.3 after 600k steps. The samples it generated is terrible under this situtation.

harryfyodor · Answer 3 · Wed Dec 30 2020 18:39:21 GMT+0800 (China Standard Time)

I did not modify any parameters. After 400K epochs, the loss oscillated around 5.3 and could not continue to converge.
Is there a more suitable parameter setting scheme to continue to optimize the loss? （ljspeech data）

Hi, i met the same problem too, have you solved it?