How to generate {train, valid}.km files?

Question

How to generate {train, valid}.km files?

Closed this issue a year ago · 5 comments

Hi everyone, congratulations on working on this model! I'm getting ready to train the contentvec model on a larger dataset. I've already made a script to generate the spk2info.dict file, which I intend to share. Do you have any tips on how to generate the {train, valid}.km, dict.km.txt files? I'm studying the code in contentvec_dataset.py but it's really hard to understand that part.

Answer 1 · 2023-07-15T21:11:29.000Z

Please refer to this https://github.com/facebookresearch/fairseq/tree/main/examples/hubert/simple_kmeans

Answer 2 · 2023-07-16T13:57:15.000Z

@auspicious3000 thanks a lot for the help! One last question: are there appropriate values for the parameters ${nshard} ${rank} and ${n_clusters}, or can I choose arbitrarily?

Answer 3 · 2023-07-16T16:16:06.000Z

${nshard} ${rank} are not the model's parameters. ${n_clusters} is an important parameter of the model. Reading the original hubert paper would be helpful to understand its importance.

Answer 4 · 2023-07-17T12:28:22.000Z

Thank you very much @auspicious3000 !

Answer 5 · 2023-08-04T06:28:49.000Z

Could you tell me how to generate the spk2info.dict file? I had generated {train, valid}.km files,but I don't know how to get it. I cannot train a new model because of the wrong spk2info.dict file.