Reduce hardcoded input and output shapes

[2martens]

In the train_AAE.py file a zsize of 32 is defined. But it is not consistently used. Sometimes the number 32 is used directly which makes it difficult to modify the code for different image sizes.

It would be helpful if you could clearly use variables designated channels, height, width and zsize (should that be different from height/width) instead of hardcoded values wherever channels, height, width or zsize are effectively used. This should be applied to the network definitions as well.

This would make it significantly easier to port your work to other datasets and/or Tensorflow/Keras.