Relations: nn_tilde's number of outlets, parameters during training, dimensions of latent space?

[CaMeLCaseError]

Hey all,

I'm just starting to train a few models using the RAVE_COLLAB jupyter notebook example. In my first attempt I was creating a "small" model and also a pretty short data set (5 hours of music). Got a 27MB sized .ts file. I didn't train prior (not sure if that matter) and exported it to get a .ts file. In puredata I have now 4 outlets on the encode and decode objects.

In my second attempt I used the "large" setting, higher fidelity (0.98) and 51 hours of dataset. It has now finished 4 million steps and I exported it to do a small listening test in realtime. I got an aproxximately 250MB large .ts file and was a bit surprised to find only 4 outlets in nn_tilde again. Other models of this size I downloaded had at least 8 or sometimes even 16 outlets. Is there any place where I can specify the amount outlets? And is there a relation between the latent space dimensions that are used during training and the final amount of outets I get? Is there a chance to gain 16 outlets from my training?

Testing my real time model that gave me the 4 outlets sounded much worse compared to the audio validation that I get during training.

Hope somebody has experience on this :)
nice regards,
Martin