Currently, I think that 1 -> 32 -> 128 -> 32 -> 128 -> 1 performs good enough and is fast enough (well... not really fast enough but its not too slow at least). See if other embedding sizes do better.
Also see if convolution size (currently 11x11) can be made smaller without loss of success. Currently the likelihood function should connect all pixels together (thanks FT), so we may not need giant receptive fields of the deeper layers.
Finally, check which optimization objective does best (final output vs all outputs. mse vs sse).
Currently, I think that 1 -> 32 -> 128 -> 32 -> 128 -> 1 performs good enough and is fast enough (well... not really fast enough but its not too slow at least). See if other embedding sizes do better.
Also see if convolution size (currently 11x11) can be made smaller without loss of success. Currently the likelihood function should connect all pixels together (thanks FT), so we may not need giant receptive fields of the deeper layers.
Finally, check which optimization objective does best (final output vs all outputs. mse vs sse).