Brainstorm ISMIR tasks

[ruohoruotsi]

Brainstorm ISMIR 2020 tasks that would be enlightened, liberated, illuminated with a proper MIDI dataset of Caribbean riddims (some with melodies and skank chord progressions)

• Release MIDI Dataset
• Use it on one or more tasks. See MIREX for more info
  • Preliminary ideas include: generative symbolic music systems, VAE latent space explorers, bassline-melody embeddings, other symbolic music information retrieval/generation tasks?

[ruohoruotsi]

This is interesting:

• UnmixDB: A Dataset for DJ-Mix Information Retrieval
• https://github.com/Ircam-RnD/unmixdb-creation

[Cortexelus]

Generation -- could try taking a pre-trained Transformer (trained on a large multi-genre midi collection) then fine-tune it on this dataset [like the LakhNES training method]

[ruohoruotsi]

Hey CJ, that's great idea ... so to summarize (1) preprocess data into trainable format (2) finetune on something like the Transformer-XL checkpoints (3) generate more basslines. What did I miss?

[Cortexelus]

That sounds about right. Convert your dataset into the format the pretrained model used. Finetune train (swap in this new dataset and continue training). Generate tons of output examples as you train.

Not sure if you've done other generative Transformer finetuning experiments, but I've done it on text with GPT2, as far as output quality, there's usually a sweet range. Too little iterations and it doesn't fit the aesthetic of your dataset yet. Too many iterations and it plagiarizes too much. The sweet range may last several thousand iterations. Every iteration you sample from has a different personality. If you only generate from your best/last iteration, the output won't be as diverse as if you were to generate from many different iterations in the sweet range.

[ruohoruotsi]

Thanks CJ, I appreciate the tips! Once I'm going on the dataset, and have some initial generated basslines. I'll share it w/ you, perhaps there's a vector of collaboration here.

I haven't fine-tuned generative Transformers. But I have built/iterated/refined lots of other audio/dsp/speech models mostly on audio-signals though.

Do you have any recommendations on MIDI python libs/frameworks that'd be useful in whipping the MIDI data into shape? I fear I may have to manually pull the "riddim"/bassline out of the songs that are multitrack, others are just riddims + drum. Also debating whether to keep the "skank" with the bassline 🙇 🤔

[Cortexelus]

cool! :D keep me in the loop hmm try pretty-midi http://craffel.github.io/pretty-midi/

…

On Sat, Jan 25, 2020 at 1:46 AM Ruoho Ruotsi ***@***.***> wrote: Thanks CJ, I appreciate the tips! Once I'm going on the dataset, and have some initial generated basslines. I'll share it w/ you, perhaps there's a vector of collaboration here. I haven't fine-tuned generative Transformers. But I have built/iterated/refined lots of other audio/dsp/speech models mostly on audio-signals though. Do you have any recommendations on MIDI python libs/frameworks that'd be useful in whipping the MIDI data into shape? I fear I may have to manually pull the "riddim"/bassline out of the songs that are multitrack, others are just riddims + drum. Also debating whether to keep the "skank" with the bassline 🙇 🤔 — You are receiving this because you commented. Reply to this email directly, view it on GitHub <#3?email_source=notifications&email_token=AASXCZN6L6L2KSZ2EKVWZV3Q7PN2TA5CNFSM4IEVUD62YY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGOEJ4WW7A#issuecomment-578382716>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AASXCZO4LLQJEGKNVC4VXK3Q7PN2TANCNFSM4IEVUD6Q> .