Need An Easy Inference Sample

[fish0510]

Hello,

I would like to run inference using your provided pretrained model without additional training or data binarization. However, when I tried to run it, I encountered many required files in the data/ directory. On the other hand, there are no clear examples explaining input information format. For instance, in style_transfer.py, in the example_run function, it does not specify the expected format of information in inp.

Could someone share your experience or provide guidance on this?

Additionally, if I want to use my own data (e.g., annotated data from OpenCpop) for inference with the pretrained model, what steps should I follow?

Thank you!

[AaronZ345]

If you want to perform customizable inference, such as in style transfer, you have two options:

1. You can replace the item_name used in inp with the information commented out in the file. For example in style_transfer.py:

# 'text_gen': ,
# 'note_gen': ,
# 'note_dur_gen': ,
# 'note_type_gen':,
# 'text_in': ,
# 'note_in': ,
# 'note_dur_in': ,
# 'note_type_in':,
# 'ref_audio': ,
# 'ph_durs':

2. Alternatively, you can refer to the GTSinger metadata format to create metadata for your data. This ensures that each WAV file possesses the required attributes. Then, use item_name to facilitate the inference process.

If you want to perform large-scale inference using OpenCPOP data, please follow GTSinger's guidelines to binarize the data. After that, run the following command:

CUDA_VISIBLE_DEVICES=$GPU python tasks/run.py --config egs/sdlm.yaml --exp_name SDLM --infer

[jackien1]

I second this as well. If TCSinger could make a Google Colab tutorial or simple inference sample that would be greatly appreciated!

[DongJiashu]

If you want to perform customizable inference, such as in style transfer, you have two options:

1. You can replace the item_name used in inp with the information commented out in the file. For example in style_transfer.py:

'text_gen': ,

'note_gen': ,

'note_dur_gen': ,

'note_type_gen':,

'text_in': ,

'note_in': ,

'note_dur_in': ,

'note_type_in':,

'ref_audio': ,

'ph_durs':

2. Alternatively, you can refer to the GTSinger metadata format to create metadata for your data. This ensures that each WAV file possesses the required attributes. Then, use item_name to facilitate the inference process.

If you want to perform large-scale inference using OpenCPOP data, please follow GTSinger's guidelines to binarize the data. After that, run the following command:

CUDA_VISIBLE_DEVICES=$GPU python tasks/run.py --config egs/sdlm.yaml --exp_name SDLM --infer

How the 2nd way knows what to infer? like option 1 you give the audio / datas so it gets input and know what to output, but im confused with the 2nd way

[AaronZ345]

If you want to perform customizable inference, such as in style transfer, you have two options:

1. You can replace the item_name used in inp with the information commented out in the file. For example in style_transfer.py:

'text_gen': ,

'note_gen': ,

'note_dur_gen': ,

'note_type_gen':,

'text_in': ,

'note_in': ,

'note_dur_in': ,

'note_type_in':,

'ref_audio': ,

'ph_durs':

2. Alternatively, you can refer to the GTSinger metadata format to create metadata for your data. This ensures that each WAV file possesses the required attributes. Then, use item_name to facilitate the inference process.

If you want to perform large-scale inference using OpenCPOP data, please follow GTSinger's guidelines to binarize the data. After that, run the following command:

CUDA_VISIBLE_DEVICES=$GPU python tasks/run.py --config egs/sdlm.yaml --exp_name SDLM --infer

How the 2nd way knows what to infer? like option 1 you give the audio / datas so it gets input and know what to output, but im confused with the 2nd way

For this way, model will find the item_name in the metadata and get all information needed.