SRD-VC
SRD-VC copied to clipboard
Speech Representation Disentanglement with Adversarial Mutual Information Learning for One-shot Voice Conversion (Interspeech 2022)
[https://drive.google.com/file/d/1HA7mlvird_xtzrd-CdqKRykFHsR4PgDp/view?usp=sharing](https://drive.google.com/file/d/1HA7mlvird_xtzrd-CdqKRykFHsR4PgDp/view?usp=sharing)
关于demo.py的问题
作者您好, 我在运行demo.py的时候,发现有from autovc.synthesis import build_model。我从autovc中复制了synthesis.py到我的目录下,但还是有错误。 Traceback (most recent call last): File "demo.py", line 140, in model = build_model().to(device) File "/data2/panl/SRD-VC-master/My_model/synthesis.py", line 22, in build_model out_channels=hparams.out_channels, AttributeError: 'HParams' object has no...
您好,我尝试follow您的工作,并迁移到其它领域,但是在训练过程中主要遇到了如下几个问题: lld_loss不收敛,导致互信息上界估计不准确,影响训练过程 使用mi_loss之后,模型参数中出现nan mi_loss随着训练过程越来越大 我尝试了调整mi_net的层数和学习率等方法,但是问题依然存在。 想请教您模型训练中的更多细节: 您的模型在训练过程中,lld_loss是否是逐渐收敛的,还是稳定在一个范围? 在mi_loss的反向传播中,mi_net的参数是否更新? mi_loss的训练过程大概如何,是否收敛?
有关MI的问题
请问你使用互信息mi的时候,有出现lld,和mi的loss为负数,以及一直不收敛的情况嘛,如果有的话,你是如何解决的呢
The link for the pre-trained weights shows that the file does not exist. Will it be possible to upload the weights in google drive? Thanks
Thank you. May I ask if you have encountered a situation where lld_loss is consistently negative (such as an abnormal value of -190)? How should this be resolved?