HSUNEH/DOSE — reverse-engineered prompt

I want this repo turned into a working local tool for drum one shot extraction from a music mix. The goal is simple, I should be able to drop one or more wav files into an input folder and get extracted one shot audio for kick, snare, or hihat saved into a results folder.

Please wire up whatever is needed so the inference flow actually works end to end with the provided scripts. Use Python 3.10, install the requirements, and make sure the model checkpoint files are expected in the checkpoints folder. I want an easy way to run it from the command line, including choosing the instrument type, input folder, output folder, and whether to use CPU or a CUDA device.

If anything is missing or unclear, check the paper page or current docs online. Also make the repo a little more friendly by adding a short run guide and sensible defaults so I can test it quickly with the sample wavs folder and see outputs in results.