วีดีโอ

I don't. The repo's license is MIT. github.com/devstermarts/PD-components/tree/main?tab=MIT-1-ov-file#readme I'd appreciate a shout out and a link to your instrument when you're done, though

No. I didn't release it. You can train it quite easily yourself, however.

@@martsm_n did you used V1 or V2? On V2 and around 6hrs dataset with 2 3060 12gb it's taking around 1 week for 1 M epochs (I'm using a batch size of 36) [do you have a discord for discussing?]

@@davidevitturini5837 This one is a V2 model with default regularization. Another V1 model has been used in this video: th-cam.com/video/LgNmYUJaSi4/w-d-xo.html

@@martsm_n thanks, really enjoyed your creativity

This is actually footage from very early and cautious interventions. As far as I can recall, it's not much more than altering the spread of latents per dimension given - in PD that'd be amplitude modulation of a signal value. You can find more videos and techniques of latent intervention in my Neural Audio playlist: th-cam.com/play/PLmNfQif1XJQ0aNdrqk8HxvwUxq90xc9al.html

Thank you. That feedback means a lot ❤

In my understanding, it's rather "sequentially plausible" in the model's understanding of what embedding should follow the one currently generated in sequence.

Can't answer that, sorry🤔

The model used in this video is trained on music I wrote in the past 20 years under various aliases and augmented from 7-8h into somewhat around 2-3 days of audio material.

so cool! 7-8 hours thats crazy long! on google GPU? I heard 2 hours is sufficient. Do you separate your songs to multitracks like bass drums, drones, snares etc to optimal the result?@@martsm_n

@@martsm_n sorry I didn’t understand “under various aliases and augumented”…

It just means I'm using all the music I have ever written under different artist names. Also, simply put, I'm sometimes augmenting the audio datasets to become bigger.@@ShihWeiChieh

My Kaggle notebook comes with instructions on training and export - maybe that helps for starters? github.com/devstermarts/Notebooks/blob/main/RAVE_Training_Template--Kaggle.ipynb I'll see to create some kind of walkthrough video at some point.

@@martsm_n thank you for this and for all the knowledge you're spreading :)

@@federicoinzerillo5333 I can recommend joining the RAVE discord channel for discussing this kind of detail. The short (and potentially unsatisfying answer) is: it all depends. Personally, I find V1 the most robust for my interest and use case. Also I achieve satisfying results with training lengths well below the recommended amount of training steps and datasets that would be considered small (1-2h). You can override potentially every setting defined in the .gin configuration files in /configs. Hope that helps.

Thank you! And no, this is all PD vanilla without additional libraries. What you're referring to are mainly arrays and graphs. You'll certainly stumble upon how these work when you dive into some PD tutorials on your way :)

@@martsm_n I’ll definitely look into it thank you!!

i got some helps from the discord and i have my first .ts file trained. I guess I have to train my prior model next for max/msp and pd usage (because the first .ts file is not compatible with my nn~ object). So i started the MSPRIOR notebook of yours then i stuck with the training cell: "ValueError: cannot reshape array of size 131072 into shape ()". Can you help me with this?

​@@ShihWeiChiehI'm not sure if here's the right place for support requests :) I'd suggest you raise this as an issue on the MSPrior GitHub repo github.com/caillonantoine/msprior

@@martsm_n ah ok, I will go there!

@@martsm_ncould you make this point here and maybe today or so? Please post a link here for Christmas. Thank you.

Yeah, they are definitely robust. I'd very much like so see/hear people's experiments with V2 and V3, however. There certainly must be advantages that simply do not fit my use case/ practices. @antoinecaillon2531 maybe?

The exact values and constellations differs from model to model. In V1 models, the first 3-4 dimensions seem to be the most sensitive and values in the +/- one-digit range lead to nice results whereas in the higher dimensions usually you need to go higher into the two-digit range to notice differences in the timbre. V2 and V3 models behave differently most probably due to the regularization equalling out things between the dimensions - I usually have a limited range +/- 2 that I can skip through without bleeding ears.