semantic segmentation support

[tkng]

We're planning support of semantic segmentation task.

Actually, internally we've already supported, but there're some remaining problems

• Current implementation does not work on FPGAs, since it breaks the channel number restriction.
• Current implementation is very slow, since using 3x3 conv for the 1st layer. We have several choice to solve the problem, but I don't know which one is the better choice:
  • replace 3x3 conv with 1x1 conv, insert 3x3 quantized conv as 2nd conv layer.
  • insert space2depth before the 1st layer.

[tkng]

fixed by #42.