Looking here, I think the sampling might always be turned off during training, since the default is sampling=False.
I think this makes everything a normal auto-encoder, since here
if sampling if False then h is returned.
I would suggest, since you're returning z (with the sampling) all the time, so you would return mu,logvar,sample (ie. z) every time. Then you can use mu for the embeddings and sample during training
Looking here, I think the sampling might always be turned off during training, since the default is sampling=False.
embedding-kit/src/embkit/models/vae/base_vae.py
Line 43 in d2c5354
I think this makes everything a normal auto-encoder, since here
embedding-kit/src/embkit/models/vae/encoder.py
Line 116 in d2c5354
if sampling if False then h is returned.
I would suggest, since you're returning z (with the sampling) all the time, so you would return mu,logvar,sample (ie. z) every time. Then you can use mu for the embeddings and sample during training