Exact command line of warmup stage

[shahaf1313]

Hey :) Can you guys please supply the command line used to train the warmup phase? I'm trying to retrain a network "from the beginning". Thanks! Shahaf

[pascal1129]

I have similar question.

The provided checkpoint (warm-up model) is 43.3 mIoU, while the reported metric in the paper is 41.6 mIoU, which quite confused me.

[panzhang0104]

You can try one of below commands to get warmup model:

python train.py --name gta2cityv2_warmupd_ls --model_name deeplabv2 --warm_up --freeze_bn --gan LS --lr 2.5e-4 --adv 0.01 --no_resume

python train.py --name gta2cityv2_warmupd_ls_S --model_name deeplabv2 --warm_up --freeze_bn --gan LS --lr 2.5e-4 --adv 0.01 --no_resume --S_pseudo_src 1

[panzhang0104]

@pascal1129 In order to make sure the released code can be reproduced, we clean the code and retrain the whole framework. So the released model is from the clean code. Since the warmup stage is trained by adversarial loss, it is unstable. The result of this stage maybe a little bit different but does not influence the result of the final stage.