Nikronic / CoarseNet

Modified version of U-net based on "Convolutional Networks for Biomedical Image Segmentation (Ronneberger et al., 2015)" paper.

Home Page:https://arxiv.org/abs/1505.04597

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

CoarseNet is not working as a low-pass filter - Halftone patterns are visible

Nikronic opened this issue · comments

Training loss cannot be declined from 30 regarding default weights.

VGG loss does not decrease at all:

...
46 , 1 loss: 9.048758506774902
47 , 1 loss: 9.113825798034668
48 , 1 loss: 8.967781066894531
49 , 1 loss: 8.963032722473145
50 , 1 loss: 8.912592887878418

Output: Full noise