Variational-Autoencoder-PyTorch icon indicating copy to clipboard operation
Variational-Autoencoder-PyTorch copied to clipboard
verified publisher icon bhpfelix

My reconstructions are bad

Open jmandivarapu1 opened this issue 5 years ago • 2 comments

Hi,

I tried running your code but even after 40 to 50 epochs my reconstructions are like below. Can you help?

Epoch_29_recon

jmandivarapu1 avatar Feb 17 '21 15:02 jmandivarapu1

+1

seolhokim avatar Jul 19 '22 06:07 seolhokim

I stumbled upon this codebase just now and reimplemented the main class here (to clear up questions about variable names and remove a lot of copy/paste in the layers. https://gist.github.com/benearnthof/6839a96a740548579a7a278cd035d622 @jmandivarapu1 To me it looks like the images were not normalized before training, I've added that to my CelebA dataset class and over the firest 5 training epochs the loss seems to steadily decrease so far. (Average loss during 5th Epoch: 1.22) I'll report back tomorrow with samples after 50+ epochs

Edit: I noticed that we optimize Binary Cross Entropy for some reason which seemed not Ideal so I swapped it for MSEloss and now my reconstructions look like this: image After 20 Epochs. Input vs Reconstructions. I'll update my gist tomorrow.

benearnthof avatar Nov 30 '23 17:11 benearnthof