TODO rescale to original resolution of 650 TODO visualize results

TODO opt batch size TODO why tensorboard lost info about tensor shapes in up components? TODO should we randomize batches TODO name scope vs var scope; need all explicit var scopes? TODO For conv transpose use kernel size divisible by stride, or (possibly better option) upsample by nearest neighbour and then do regular convolution TODO bias in conv nets - with/without batch norm TODO how to reuse variables?