search

wasidennis / AdaptSegNet

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

About the training of AdaptSegNet(GTA5 -> Cityscapes) and usage of Synscpaes #98

Open LinfengYuan1997 opened 3 years ago

LinfengYuan1997 commented 3 years ago

Thanks for your fantastic sharing code!

I have 2 questions about your code:

  1. Recently I tried to train your code AdaptSegNet(GTA5 -> Cityscpaes multi ResNet) via PyTorch 1.4.

The only modifications I made are the argument 'align_corners = True' in nn.Upsample() and small bugs while recording and printing the loss using PyTorch1.x.

And I have trained the model in 120K iterations but only got the highest mIOU on Cityscapes about 40.x% at 40K iteration.

image

I am wondering thether it is normal because GAN is unstable, or there are some problems during my training.

And here are some warnings when runnig the training code: image Is there any problem about the optimizer?

  1. I plan to use Synscapes to train and validation the UDA of segmentation. I am wondering your setting while using the Synscapes (the train val splits and any hyperparameters tuning while training)

Looking forward to your reply~

LinfengYuan1997 commented 3 years ago

Maybe all the Synscapes pics will be entered for training because we don't use them while validating on Cityscapes?

Any hyperparameters tuning by the way?