Create a conv autoencoder which has more conv layers

[kingofoz]

The current one soumith wrote based on unsup package has only one conv layer for encoder and one conv layer for decoder. I tried to add more conv layers to it, but then it is very difficult to train. Can anybody write such a demo which has more conv layers?

[kingofoz]

By the way, I am using the conv-psd part of the demo soumith wrote.