yuangan / EAT_code

Official code for ICCV 2023 paper: "Efficient Emotional Adaptation for Audio-Driven Talking-Head Generation".
Other
269 stars 30 forks source link

关于预训练使用的权重 #35

Closed Mhq-we closed 3 months ago

Mhq-we commented 3 months ago

作者您好,非常感谢您的分享!请问./ckpt/pretrain_new_274.pth.tar这个权重是如何获取的?是按照pretrain的step1从头开始训练274个epoch吗?

yuangan commented 3 months ago

您好,感谢关注,我们的pretrain_new_274.pth.tar是通过微调这个开源的OSFV模型得到的。具体改进可以参考论文的Sec.3.1.1

Mhq-we commented 3 months ago

感谢回复!请问微调OSFV模型所用的数据是您过滤后的Voxceleb2嘛

yuangan commented 3 months ago

是的,用的数据是voxceleb2和mead