Exp12: Overfit Test
Micky774 opened this issue · comments
Try to overfit as best we can with beta = 0 to sanity check whether we have a sufficiently complex network.
Currently running a 90 epoch overfit test. Results pending
Initial conclusions: brightness had a large effect on quality, as brighter regions tended to be better reconstructed. I believe skip connections may solve this. Normalization of input or layers may also help. Furthermore, a noticed deficiency in the length of the FC layer has been fixed as a result of this overfit test. A new overfit test will be run to confirm the significance of the changes