I have implemeted a Variational Autoencoder(VAE) with a prior different from unit gaussian. This gives me extremely sharp reconstructions compared to normal VAE(small values for reconstruction loss). But the generated samples are of extremely poor quality. Why is that? Also would the generated samples improve if trained longer. Right now I'm training for 50 epochs.

Edit: The prior is similar to Laplace distribution. I'm working on Mnist dataset. Currently I'm comparing the reconstruction loss of vanilla vae with this custom vae for validation. Also the reconstructed images have no blurring. But it seems the latents learn nothing. I tried training for 200 epochs with no difference in generated samples but reconstructions got better.

  • $\begingroup$ Welcome to CV! Your current question is too broad I'm afraid. Different from unit Guassian how? What kind of images; how many observations; how are you validating? Without additional information and perhaps some results it is hardly possible to answer your question. $\endgroup$ – Frans Rodenburg Aug 29 '18 at 8:21
  • 1
    $\begingroup$ To help us help you, could you please show: 1) a few outputs of your full VAE (encoder + decoder), with the corresponding training inputs? 2) a few samples from the VAE (i.e., outputs from decoder only, when fed with sampled latents); 3) & 4) like 1) & 2), but for your baseline VAE? If I understand you correctly, the output of the full VAE (encoder + decoder) on the training images is great (very similar to inputs), but when you start sampling latents and pass them the through the decoder, you get crap. If so, the answer is very simple - but please add those images! $\endgroup$ – DeltaIV Aug 29 '18 at 9:28

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.