Closed HUSTLX closed 4 years ago
When we use the style vec as additional input, we found using 2e-4 would lead the network to diverge… So we lower the learning rate to … to 2e-5 and then to 1.5e-5… Experiments show 1.5e-5 gets better results.
On Apr 11, 2020, at 10:44 AM, HUSTLX notifications@github.com wrote:
usually the learnling rate is 0.0002,why do you use 0.000015?is there a clear experimental proof?
— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub https://github.com/yiranran/Unpaired-Portrait-Drawing/issues/4, or unsubscribe https://github.com/notifications/unsubscribe-auth/AD76RAEQSMBIDDV5L4ADHETRL7KQXANCNFSM4MF2I6MQ.
usually the learnling rate is 0.0002,why do you use 0.000015?is there a clear experimental proof?