search

acids-ircam / RAVE

Official implementation of the RAVE model: a Realtime Audio Variational autoEncoder
Other
1.35k stars 184 forks source link

Number of dimensions of exported model does not agree with Tensorboard statistics #267

Closed leoauri closed 11 months ago

leoauri commented 11 months ago

Hi, I am training a model with RAVE using --config v2 --config causal. In tensorboard I can see, for example, that fidelity 0.8 is reached with 7 dimensions. However if I export the model with --streaming --fidelity 0.8 and load it in ~nn model_name decode, the component has 8 inlets. In fact all fideltiy settings I have tried result in an ~nn component with some number n^2 inlets, always a power of 2 and not agreeing with what I see in Tensorboard.

Expected: number of inlets on ~nn component agrees with the number of dimensions needed for that fidelity as reported during training.

Am I missing something here? Thanks, L

domkirke commented 11 months ago

Yes, for export dimensions have to be a power of 2 for optimization reasons, such that the number of dimensions it is rounded up to the next power of 2.