demo for facebook musicgen

features

• run in docker
• can continue itself as much as vram you have
• almost zero conf

known bugs

• output duplicates to output.wav
• lame code
• need to loadout generated parts to cpu, or may be even disk, for really long generations
• there seems problem in audiocraft library last version, progress shows wrong total,

Installation:

• clone this repo
• install nvidia/stuff
• install docker
• install cog

  sudo curl -o /usr/local/bin/cog -L "https://github.com/replicate/cog/releases/latest/download/cog_$(uname -s)_$(uname -m)"
  sudo chmod +x /usr/local/bin/cog
  

I think it will work in ~/bin/ too if PATH configured, but I not tried

usage

(all run in repo directory)

default prompt, random seed, 10s duration

cog predict

set prompt seed and duration:

RUN

 cog predict  -i prompt='C major, Roland TR-808, Korg KR-33,  celtic dance, violin, moderate, jazz, disco, rythmic, complex melody' -i duration=90 -i seed=1234 -i cfg_coef=3 -i temperature=1 

run with -i burn_times=n to burn first part with chroma n times, and use chroma of result 1st part as chroma to all subsequent "continues"

---

config

to change model size and other params, edit musicgen_config.json (i have .example)

currently working model sizes:

• small
• medium
• melody
• large

output wavs in ./out/ directory

tested on windows+wsl+docker rtx3070 and linux+docker rtx4090