backward error in training

[yangli-lab]

Hello, I got an error when try to train on my own dataset, in the optimizer.step() . Do you have any idea about it?

[ZhenglinZhou]

Hi @yangli-lab,

What is the number of facial landmark in your own dataset? It seems like that the number of facial landmark has been set incorrectly. Remember to modify the config file.

[yangli-lab]

The landmark number is 98, and I set the data_definition in alignment.py as 'WFLW', as you can see below

[yangli-lab]

more detailed error logs

[evemtics]

更详细的错误日志

你好请问解决了这个问题吗，我也存在这个问题

[yangli-lab]

The issue arose due to the default loading of optimizer parameters corresponding to those used by the authors during pretraining, which resulted in errors when attempting to modify optimizer-related parameters during training. I resolved this problem by explicitly setting up a new optimizer. Hope this will be helpful