Small improvements that could be made

[axeld5]

Hello, I am currently running moshi-finetune for a personal project (goal : checking if small amount of QA samples can enable Moshi to learn latently knowledge about a topic)

After making it work, a few notices:

• It would be better to more emphasize the need for stereo datasets : it is currently a very small note in the repo, but is necessary to make it work ; perhaps a few additional words in the tutorial could be helpful ; perhaps even add a "check if ready" python script that verifies whether or not the dataset is completely ready? Use soundfile, check if all audios have info.channels = 2 (and if there is also a need for the frequency to be 44.1k hz, relabels the frequency)
• Having to rm -rf the run_dir after each bugged attempt of launching the train script was tedious : could be interesting to have in command line some sort of "-r" in args that deletes the folder if present? Did I miss it?
• Having to set-up export Cuda Visible Devices = 0 and Cuda Device Order = PCI_BUS_ID every time I relaunched my VM was also tedious : is there something I missed here? Or perhaps there is a way to set it up easier?

If you think it's interesting, I have also no problem contributing to a PR that could improve all said elements!

[hippolytepilchen]

Hello,

Thank you for your feedback. We have added an argument that allows you to overwrite an existing checkpoint directory, which should make re-running experiments easier.

Regarding the environment variable export issue, I think you could include those export commands directly in the .sh script you use with sbatch. That way, they’ll be automatically set each time the job runs.