[Help]: Format of CustomSVCDataset

[open-mmlab]

[Help]: Format of CustomSVCDataset

[RMSnow]

Hi @mysxs, thanks for your attention!

It seems that your provided train.json is not generated by Amphion's code. Our instruction of CustomSVCDataset means that you can prepare your wavs as the following format:

[Your Custom Dataset Path]
 ┣ singer1
 ┃ ┣ song1
 ┃ ┃ ┣ utterance1.wav
 ┃ ┃ ┣ utterance2.wav
 ┃ ┃ ┣ ...
 ┃ ┣ song2
 ┃ ┣ ...
 ┣ singer2
 ┣ ...

After that, you need to specify your custom dataset name and its path ([Your Custom Dataset Path]) in the exp_config.json:

    "dataset": [
        "[Exisiting Dataset Name]",
        //...
        "[Your Custom Dataset Name]"
    ],
    "dataset_path": {
        "[Exisiting Dataset Name]": "[Exisiting Dataset Path]",
        //...
        "[Your Custom Dataset Name]": "[Your Custom Dataset Path]"
    },
    "use_custom_dataset": [
        "[Your Custom Dataset Name]"
    ],

After that, when you preprocess the custom dataset (--stage 1), the Amphion will automatically produce the train.json and test.json for you (see this line), where the generated json files will have the Duration key.

You can follow the steps above. If there are still any bugs, feel free to let us know.

[mysxs]

Hi RMSnow, thanks for your reply!
okok, it turns out that train.json is automatically generated by Amphion. I just found that the directory of my data set is one level higher than CustomSVCDataset. I will modify it and try again~

[mysxs]

Hi RMSnow, thanks for your attention!
I reorganized my dataset according to the CustomSVCDataset format. I added two pictures to illustrate my dataset, but the previous error was still reported. Amphion did not automatically generate train.json.
the format of my dataset:

exp_config.json:

error:

[viewfinder-annn]

Hi @mysxs !

Seeing your provided screenshots, I came up with two possible reasons:

1. You are not specifying use_custom_dataset field in your config.json file. You need to add a field like this if you havn't done it:

"dataset_path": [
    "ESD": ...
]
"use_custom_dataset": [
    "ESD"
],

2. You are setting the processed_dir same as your raw audio dir. It's recommended not to do so. During the preprocess stage a dir with the name of your dataset name will be created in the processed_dir path. In this case it will create another "/egs/ESD" dir which may cause overide of your original files. You can try to set this field default("data").

[mysxs]

Hi @RMSnow @viewfinder-annn,thank you for your response. Following your advice, I corrected my dataset path, and I successfully resolved the issue. Thank you!