Retrain and finetune AudioNet

[niladridutt]

While training the network (l2face_trainer), are we fine tuning the audio net? Or is the pretrained weights good enough to be used on any video?

[niladridutt]

@zhangzjn @YongLiuZJU Any information on how to train the Audio net will be greatly appreciated

[zhangzjn]

We recommend fine-tuning the weight on the custom dataset, even if the weight has some generalization capability.

[niladridutt]

Thanks, but in the code I can't see any method to optimize the audio net (netA) here. I can only see the face re-enactor generator and discriminator weights getting updated

[zhangzjn]

You need to modify the code a little bit about the audionet if you want to train on the custom dataset. This link may help.

[niladridutt]

I see. There seems to be some differences between the 2 codes - addition of a latent vector and some changes in the net architecture as well. Is it possible for you to push the code for training the audionet if posssible? It would be a huge help. Thanks.

[zhangzjn]

Sorry, this code will be considered for update at a later date.