Retrieval-based-Voice-Conversion-WebUI 有没有使用mac进行安装的小伙伴？训练原声的时候显示成功了，但是特征检索库是空的是什么情况？log下没有那个index的文件。

有没有使用mac进行安装的小伙伴？训练原声的时候显示成功了，但是特征检索库是空的是什么情况？log下没有那个index的文件。

Open Kar1chan opened this issue 1 year ago • 5 comments

你们能成功吗？特征库是空的 97fb60f77711616a4392e237e2d627d9 推理音色是有的 eb6b92a1dffef34a91ea64a7a3c226c7

Jul 06 '23 02:07 Kar1chan

用Mac安装了吗我也想问😊

Jul 08 '23 18:07 Nemophilister

同問，能成功推理而效果不錯，但是不知道沒有這個影不影響。

Jul 08 '23 21:07 Naozumi520

I got the same issue here, but model inference can be performed successfully after some tweaks. One needs to indicate PYTORCH_ENABLE_MPS_FALLBACK=1 (e.g., add os.environ["PYTORCH_ENABLE_MPS_FALLBACK"] = "1" at the top of infer-web.py) to prevent NotImplementedError: The operator 'aten::_weight_norm_interface' is not currently implemented for the MPS device.. Otherwise, this aten::_weight_norm_interface error can lead to AttributeError: 'NoneType' object has no attribute 'dtype somehow.

commit: 3dbba6ae7471a3fabaa42f18d137e50ec04cf879 OS: MacOS Ventura 13.4 Hardware: Apple M1 Max

Jul 09 '23 20:07 yayuanli

@yayuanli You're right. Some functions are not supported by MPS with PYTORCH_ENABLE_MPS_FALLBACK it can be done on the CPU. But the issue here is the .index file is not being generated during training. If you click 训练特征索引, you will noticed the script crash. Do we know 训练特征索引 is not currently supported.

Jul 10 '23 05:07 Naozumi520

@Kar1chan Try to downgrade faiss-cpu to 1.7.0.

Jul 12 '23 06:07 Naozumi520