bingoHong

thanks to apply the dropout trick. i expect the CUDA accelerated version.

this answer debug my not square image, thanks!

i found the same problem there is no moving_average process. how is to compute gn in test step?

@shenekeng853 i add the ema process, maybe you can try it https://github.com/Bingohong/GroupNormalization-tensorflow/blob/master/README.md

参考TimeSeriesDataGenerator，可以写个自己的

on window sys, we can run _python yad2k.py yolo.cfg yolo.weights model_data/yolo.h5_ to generate the h5 file

- remove the yad2k document from "C://user//", - put it in "C://root" location. And run - cd yad2k - python yad2k.py yolo.cfg yolo.weights model_data/yolo.h5 it works ! you can try

i train u-net. utilize 3 different normalization method: - batch normalization by keras in-build layers - group normalization without moving average by other github - group normalization with moving average...

i upload 3 method result log file in compare_log dir. you can look out these. however, i think the gn without moving average get a lower val loss about 0.2....

Why this happen?