search

wasidennis / AdaptSegNet

Learning to Adapt Structured Output Space for Semantic Segmentation, CVPR 2018 (spotlight)
847 stars 205 forks source link

Can not reproduce the result of VGG net #28

Open EthanZhangYi opened 5 years ago

EthanZhangYi commented 5 years ago

I directly dowloaded this repo and trained the model. The results of Res101-single-level and Res101-multi-level are reproduced, even that the training is not stable. However I can not reproduce the result of VGG16-single-level.

I trained it with the same hyper-parameters of Res101-single-level, and the best result is 32.56%. But the result in the paper is 35.00%

Do I need to use another group of hyper-parameters?

wasidennis commented 5 years ago

Hi @EthanZhangYi, for the VGG16-single-level model, we do still use the same \lambda = 0.001 and other hyperparameters as in the ResNet-101 model, but the convergence is slower that we obtained 35% after training for 130k iterations.

chentao2016 commented 5 years ago

Hi,could you share the training code of VGG-single-level.I am interested in that but I get poor result.I have problem with changing the code for VGG training.

HqWei commented 5 years ago

@EthanZhangYi Have you reproduced the source-only result? I use VGG with an image size of 1280x720. When testing in the target domain, the image size is resize to 1024x512. But the source-only MIOU is only 15.

EthanZhangYi commented 5 years ago

@HqWei I have NOT trained the source-only model.

HqWei commented 5 years ago

@EthanZhangYi Have you reproduced the source-only based on resnet101(PSPnet, deeplabV2 or deeplabV3)? I have trained on Cityscapes and valuated on GTA5, I get 30.28(MIOU). However, when I trained on GTA5 and valuated on Cityscapes, I get around 10. I'm confused.

EthanZhangYi commented 5 years ago

@HqWei I have not trained this kind of model. But I think ~10% must be wrong.

wasidennis commented 5 years ago

Btw, we use the VGG pre-trained model on ImageNet: https://www.dropbox.com/s/r8dx8xuqpsr3do3/vgg16-00b39a1b-updated.pth?dl=0.