The hidden_states during prediction stage is the same when a custom dataset is used

Doubiiu / CodeTalker

[CVPR 2023] CodeTalker: Speech-Driven 3D Facial Animation with Discrete Motion Prior

MIT License

515 stars 57 forks source link

The hidden_states during prediction stage is the same when a custom dataset is used #62

Open Shirley-0708 opened 10 months ago

Shirley-0708 commented 10 months ago

@Doubiiu When I use a custom dataset, hidden_states of the audio encoder in each frame are the same during the prediction stage. More details can be seen from #issue-1394397224. This problem has been bothering me for a long time, sincerely hope that you can give me some advice!

paopao1027 commented 6 months ago

hello,how to get the such files like fdd.txts and lve.txts for different dataset?

Doubiiu commented 5 months ago

hello,how to get the such files like fdd.txts and lve.txts for different dataset?

see #64 .

Jackson-coder commented 5 months ago

Did you solve the problem? I also face the same situation.