search

Doubiiu / CodeTalker

[CVPR 2023] CodeTalker: Speech-Driven 3D Facial Animation with Discrete Motion Prior
MIT License
515 stars 57 forks source link

关于在vocaset上训练meshtalk问题 #67

Closed AlongLove closed 8 months ago

AlongLove commented 8 months ago

有在meshtalk的Issues里有看到您提问的问题，关于训练不收敛的问题，请问您的解决方式是什么呢？还请不吝赐教，谢谢！

Doubiiu commented 8 months ago

你好，我是通过调节超参，例如batch size，learning rate等使得模型训练可以收敛。具体的数值已经不是很记得了。

AlongLove commented 8 months ago

好的，非常感谢，它里面设置的segment_length在step1是32，step2是64，这个有调节过么？

Doubiiu commented 8 months ago

也有的，我好像大概试了8的倍数，大概16到48这样

AlongLove commented 8 months ago

好的，感谢赐教

Doubiiu commented 8 months ago

OK I will close this issue then. Please reopen it whenever you have any other questions.