sksq96/pytorch-vae

A CNN Variational Autoencoder (CNN-VAE) implemented in PyTorch

A CNN Variational Autoencoder in PyTorch

Implemetation of David Ha's research on Word Model.

pytorch-vae's Issues

what is the size of input image?

mean and variance

Name 'vae' is not defined

Hello, I tried vae-cnn.ipynb file but seemingly vae object is never defined but your code is using it. Can you explain this? Also, there is FileNotFoundError: [Errno 2] No such file or directory: 'vae.torch' error.

VAE

Hello,

How can I look at the latent space in the pytorch VAE code? I know how to do this in keras, but have not yet figured it out in pytorch.

Thanks!

Give credit to World Models

Hey thanks for having this code up.

However, I have realized that it is almost completely identical to David Ha's World Models research https://arxiv.org/pdf/1803.10122.pdf and code including the code here: https://github.com/ctallec/world-models

Please give credit where credit is due.

loss function?

I am not sure if I am correct. It seems that the loss function is directly a copy from the pytorch's official implement of VAE. But it seems to be the opposite number of the loss in the function.

In the paper the loss should be -DKL+log p(x|z)
and here are we minimizing DKL - log p(x|z)?

I am not sure if I understand correctly about F.binary_cross_entropy but from the document of nn.BCELoss torch.nn.functional.binary_cross_entropy it should be returning -log p(x|z).....

Data normalization

I wonder is there any effect due to normalized data ?
I don't see you normalized your data.