VAE and VAEGAN producing the same image. Mode Collapse?

Deep Learning
1

I have tried to implement the standard VAE and the CVAEGAN model (without the class labels) on the frames of the atari game breakout. But for some reason the model always outputs an image which looks like the average of the frames in the training set. I found that this is possibly the mode collapse problem but I am unsure as to how to rectify this.
Any ideas?

2

Can you provide snippet of the code or the github link to the code?

3

In the models folder, you will find the cvaegan model and the vae model. In the training folder are the training scripts for them.

You can also find pytorch trained model weights in the Training folder.

Thanks!