grace9994 / comosvc Goto Github PK

If not comment out the line os.environ["CUDA_VISIBLE_DEVICES"]='1' in train.py, training will encounter an error: RuntimeError: No CUDA GPUs are available.

i have a little errors when i launch :
python preprocessing2_flist.py
67%|███████████████████████████████████████████████████████▎ | 2/3 [00:00<00:00, 282.36it/s]
Traceback (most recent call last):
File "preprocessing2_flist.py", line 46, in
for file_name in os.listdir(os.path.join(args.source_dir, speaker)):
NotADirectoryError: [WinError 267] Nom de répertoire non valide: './dataset\the_prprocessed_data'

and there a empty file call the_prprocessed_data need to be dataset folder, but need to copy on \dataset\YOURNAME\

感谢工作！之前评估so-vits-svc发现有音色泄露问题，因此该项目加入了聚类，特征检索等许多办法来解决这个问题，我看了您的展示音频，没有发现音色泄露现象，我也查看您的论文也没看到提到音色泄露问题。请问，您在CoMoSVC实验过程中有没有发现音色泄露的问题，以及是如何解决的? 期待您的回复 。

This repository seems interesting but I can't execute any code as I don't have the appropriate hardware. Could there be a Colab notebook coming out with this?

Thanks for your job. I have trained comosvc on vctk-80people dataset for 16w steps, and I inferred on 16w steps frozen weights, but the vc out results is not good as expected. Any advice to improve the vc results? Looking forward to the reply. @zhenye234 @Grace9994

按照论文的结果，扩散模型和一致性模型的推理速度应该有较大差异，但是我训练出的模型，推理速度并没有太大的差异

为什么学生模型比教师模型大三倍，能否减小？

我在网上找不到任何有关CoMoSVC的描述，也无法在自己的设备上复现该模型

I observed that training loss is decreasing in overall, but validation loss is increasing. Have you experienced the same behavior?

Thanks.

请问训练consistency model时的配置文件应该是什么样的呢？直接copy 训练teacher model时的配置文件吗？

你好，非常感谢您能开源如此优秀的论文代码，请问论文里的客观评价指标的测试集是怎样构成的呀，并且他们的代码是怎样的

No such file or directory: './configs/diffusion.yaml'