Even though the reconstruction loss has (seemingly) decreased to near 0 with a generic VAE (aka one that is densely connected), the reconstructed sequences are not anywhere close to the target sequence. This suggests that the latent (z) variable is not being used to learn a proper encoding of the training set. Related to #92.
Even though the reconstruction loss has (seemingly) decreased to near 0 with a generic VAE (aka one that is densely connected), the reconstructed sequences are not anywhere close to the target sequence. This suggests that the latent (
z) variable is not being used to learn a proper encoding of the training set. Related to #92.