spleeter icon indicating copy to clipboard operation
spleeter copied to clipboard

[Discussion] Training documentation?

Open GelflingHand opened this issue 4 years ago • 2 comments

Can we get some more documentation on training? There's a small section in the wiki, but there's no explanation of the config file. What do all the options mean (in layperson's terms, to the extent possible), and what are the ranges of acceptable values?

GelflingHand avatar Feb 27 '21 08:02 GelflingHand

Hi @GelflingHand, Parameters of the config files are described in this wiki page. There may be some information missing, so feel free to ask, we'll do our best for inserting missing information in the wiki if needed.

romi1502 avatar Mar 01 '21 08:03 romi1502

Ah, thanks! Not sure how I missed that. More detail on what the parameters mean would be useful, though. Eg what's a frequency bin? Is it affected by frame_length, frame_step, and T? How does changing the various Fourier transform parameters affect the output? n_chunks_per_song - how does changing the number of chunks affect the output? Does it make sense to use more chunks for longer songs and fewer chunks for shorter songs? What difference does changing the learning_rate make, and what's the range of acceptable values? What does changing batch_size do? How does Bi-LSTM differ from U-net, and how would you specify the former in config? (ie what's the equivalent of 'unet.unet', and are different params needed?)

I realise some of this might not be easy to explain to people who don't understand the math, but whatever you can manage will be appreciated.

GelflingHand avatar Mar 01 '21 22:03 GelflingHand