An implement of Unet used for competition Jinnan2!
Arxiv PDF: https://arxiv.org/pdf/1505.04597.pdf Here!
orignal architecture:
we modify the output and make all Conv padding as "same" ,thus remove the Crop Layer.
ours architecture will be seen at ./imgs/Unet_ours.png
CUDA : 8.0
cuDnn: 6.0.21
tensorflow-gpu : 1.4
keras : 2.2.4
I am so sorry that i can not try my implementation on competition(actrually we use another model), but cause of interest, I'll finish the rest of the code.