search

facebookresearch / demucs

Code for the paper Hybrid Spectrogram and Waveform Source Separation
MIT License
8.35k stars 1.06k forks source link

Guide to Training a new model? #410

Open thelatebphelium opened 1 year ago

thelatebphelium commented 1 year ago

❓ Questions

heylo! first, i want to say many, many thanx to Alexandre and everyone else who has contributed to this work!!! incredible!!!!

i am trying to demucs some noisy mid-90's tunes from my band. they are old 4-track cassette mixes that are really a ball of noise. :) Demucs does a pretty good considering how fuzzy they are. but i am finding that it is having trouble finding the bass guitar. Demucs4 is even worse.

i want to train my own model for these recordings. i still have some of the 4-track sessions. so i could feed it individual parts for training.

how should i go about making a dataset and then start training a new model?

i've seen a couple of threads, but i'm still unclear on how to make a dataset or what should i include in a dataset.

thanks!

lucellent commented 1 year ago

This is the training guide https://github.com/facebookresearch/demucs/blob/main/docs/training.md

But I'm interested in it too, some parts are unclear to me and I don't think I'd manage to train a model on my own

GeradeHouse commented 1 year ago

I have been trying to set-up a script to replicate the training of the demucs_ht model, but the instructions of the training guide are not clear enough for me. Is there anyone who is willing to share his/her code how to train the demucs_ht model on the [Musdb HQ] dataset ?