yzwxx / vae-celeba Goto Github PK

HI,

Thank you for sharing your code.

However, when I run the plain VAE code using the default parameters, I cannot reproduce the result you post on the REAMED page.

The reconstruction image and random image I generated using your code is:

Can you provide some suggestions on how to reproduce your result?

I'm training the model with different datasets and image sizes. I have noticed that if I use images larger than 64x64, all the losses become NaN after a certain point. It seems like it happens earlier as the image size increase (for instance, it happens after 77 epochs with 128x128 images and after 3 epochs with 1024x1024 images). Do you happen to know why this is happening and do you have any advice to address it?

Thanks!

Why does the implementation differ from the paper? Is this per intention? The file model_vae.py for example has a final filter depth of 512, and uses a 5x5 kernel size. What produces the best results? The paper implementation or the code?

I understand that the author may not have time to maintain the code.

For those who use the code, please be aware that the param train_size is not effective. It is only used for logging purpose.

If I want to use the encoder only as a cnn, what layer is the one containing the embedded coding that I need to use before the soft max? that is, I want to use the encoder with the trained parameters as a pretrained CNN, But I am not sure what will be the layer that has to be connected to the softmax. This one is the equivalent the last layer of a CNN before the softmax layer, correct?

Thanks for the good code!
I have two questions:

1. Is UpSampling2dLayer+Conv2d better than DeConv2d? How about the results of DeConv2d?
2. How about the results of using the original VAE loss without perceptual loss?

Can you provide an example on how to reconstruct and generate new samples after training the network?