when I train pwcnet using FlyingChairs with train.prototxt, It doesn't converge.
I modified train.prototxt :
flyingchairs lmdb and split.list directory
crop size : 320x448, 384x448 tried both crop size, both results were similar
batch size : 4 (used 2 gpu, then effective batch size is 8)
after train, I tested the model on FlyingChairs test set (labeled as 2 in test_train_split.list)
EPE was 10.9164379646
anyone can help me?? any idea you give will be thankful!!
when I train pwcnet using FlyingChairs with train.prototxt, It doesn't converge. I modified train.prototxt :
after train, I tested the model on FlyingChairs test set (labeled as 2 in test_train_split.list) EPE was 10.9164379646 anyone can help me?? any idea you give will be thankful!!