AndreevP
commented
3 months ago Hi! Sorry for late reply, I have missed this issue:(
Unfortunately, the source code for training cannot be made available due company policy.
`
As a part of this submission supplementary material, we provide all source codes that are needed to train and infer our models. We also attach configuration files that contain all the necessary information regarding the model’s specification and hyperparameters." in Appendix D
` It was provided as part of the submission for reviewers only, not for public release.
However, training hyperparameters can be made available and that is a mistake that configs do not contain this info. Apologies for that. If you still need this info, I attach configs with training hyperparameters. configs.zip
I am sorry that I cannot provide more:(
My advice is that you should take the HiFi++ architecture from this repo and use original HiFi-GAN codes for training. This should decently reproduce our work.
If you need any other advice, please reach me by email andeev.pk.personal@gmail.com
Again sorry:(
Respectfully, Pavel Andreev
Dear Paveel,
Thanks a lot for releasing the supplementary material at https://arxiv.org/pdf/2203.13086v4.pdf and releasing the architecture code at this repo.
However, I don't find in the Github repository any hyperparameters used for HiFi++ training nor proper training code (the trainer is actually an inference code), although the ICASSP 2023 paper states "All training hyper-parameters and implementation details will be released with source codes." on page 3, and "As a part of this submission supplementary material, we provide all source codes that are needed to train and infer our models. We also attach configuration files that contain all the necessary information regarding the model’s specification and hyperparameters." in Appendix D.
Perhaps I've missed something, but the yaml files in https://github.com/SamsungLabs/hifi_plusplus/tree/main/configs are only for inference using the pre-trained HiFi++ control points found in release version 1.0.
As we would like to test HiFi++ with other datasets and bandwidth degradations than those you used in your article, can you give me an expected date for the training code and hyperparameters to be made available, so that we can compare HiFi++ for research purposes in the context of a publication under writing citing your work?
Best regards,