Hello,
A message from @demerle.
A few weeks became a few months but I finally pushed after v1.1.0 to the acids-ircam repository. Along many modifications to the architecture and the training to make it more stable, I also pushed two max4Live devices designed by @esling and myself. The devices feature a small trick to make the n-dimensional timbre latent space into a 2D map that you can use as a “coarse” latent explorer (while the actual latents are still exposed for fine control). I added checkpoints that work with the M4L devices here. Let me know if you encounter any issue, and more importantly let me know if you train a model and it sounds good and/or you have fun with it Little changelog for 1.1.0 :
- Max4Live Devices with latent map,
- Export scripts now create the latent map + automatically picks last checkpoint,
- Moved to local attention in the transformer layers, which should result in much more stable streaming after export,
- Inference notebooks can take .ts files,
- Change regularisation in the autoencoder to avoid latent collapse,
