Learned latent dimension changes from training phase 1 to 2

[moiseshorta]

Hi!

I've been successfully training a new RAVEv2 model.

What I noticed is that when training a model, the learned latent dimension changes radically from training phase 1 to 2.

I noticed a change of 32 dimensions in the first training phase to just 1. Is this expected behavior for version 2?

I was training using the default v2 config, with the default ELBO regularizarion, using lazy dataset callback.

Thanks!

[gyuchulm]

Did you install and run it with the default setup? I've no issues with train and export, but there are no sounds, comes up in max and Supercollider(tested in rave v1). If you use different set up, Is there any tips for training?

[moiseshorta]

@gyuchulm Yes, this is all trained on the default setup for v2. I also experienced the same, it did not work when exporting and running on nn~ even tho I updated my pytorch version...

[domkirke]

What kind of data was this trained on? Did you provide additional configurations?