Training model with a dataset

[gcunhase]

Any hints on how to use NB2 to train a dataset (say a directory with multiple audio files) and then use the trained model to generate one of those samples?

Thank you in advance

[fatchord]

@gcunhase Yeah it should be easy enough - you could try Pytorch's Dataset class. Whether the model will produce one of the samples is a gamble since the model would be unconditioned and you wouldn't have much, if any, control over what it outputs.

[gcunhase]

@fatchord Thank you for your reply, would it be possible to condition it somehow?

[fatchord]

@gcunhase Have a look at notebooks 4a and 4b - there you'll see how I do it.

[hyzhan]

Is that have some sample audio？ I am trying to train on the LJspeech dataset for seq_len=8000,batch=12,learning_rate=1e-3 warm up 4k step and then decay, my loss is around 1.7

[fatchord]

@hyzhan what model are you training? By the way the seq_len should be around 1000 for more efficient training - also make sure it's cleanly divisible by your sprectrogram hop length.

[hyzhan]

Thanks for reply, I have successfully trained in the alternative model, it sounds good, thank you very much. But the strange thing is that the alternative model can converge in a dataset of about 5 hours, but the dataset is often jittery in 30 hours, causing the latter to generate audio with loud noise.


Above are two large data sets, the smaller ones below:

Could you please give me some advice?