Results are not consistent with the reported

ZJULearning / MaxSquareLoss

Code for "Domain Adaptation for Semantic Segmentation with Maximum Squares Loss" in PyTorch.

MIT License

113 stars 22 forks source link

Results are not consistent with the reported #8

Closed omerlandau closed 4 years ago

omerlandau commented 4 years ago

hello, i've been training the GTA5 model couple of times, even when using pretrained pth on source only, maximum miou is 43.6. using the instructions gave in this repo on the IW maximum + multi. how come you've managed to score above 46 training this way, this is a big gap. thanks in advance, omer.

greeneggsandyaml commented 4 years ago

Hi @omerlandau, did you manage to solve this issue? I am also having difficulty reproducing the results.