search

wasidennis / AdaptSegNet

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

reproduce the baseline (ResNet) performance #81

Open seominseok0429 opened 4 years ago

seominseok0429 commented 4 years ago

I want to reproduce the baseline (ResNet) performance of table 1 in your paper(AdaptSegNet), can you share the same parameters as iteration? please.

For example, in the code, iteration is 250,000, but the amount of GTA5 dataset is about 25000. So, is it really true that about 10 epochs with only 10 learning per image?

adricarda commented 4 years ago

Were you able to reproduce the baseline results? I tried with the same hyper-parameters indicated in the supplementary material but sill no luck. I can reach at most 25% in terms of mIoU on the Cityscapes validation set. Can you share the weights maybe? thanks.

chinacui commented 4 years ago

@adricarda Same question. Have you solved it successfully? I have trained the Seg model source only and got ~24-25 mIoU accuracy. Thanks.

seominseok0429 commented 4 years ago

Were you able to reproduce the baseline results? I tried with the same hyper-parameters indicated in the supplementary material but sill no luck. I can reach at most 25% in terms of mIoU on the Cityscapes validation set. Can you share the weights maybe? thanks.

The hyperparameters were all the same. But we run 12 epochs. It is like 30000 iterations.

seominseok0429 commented 4 years ago

@adricarda Same question. Have you solved it successfully? I have trained the Seg model source only and got ~24-25 mIoU accuracy. Thanks.

The hyperparameters were all the same. But we run 12 epochs. It is like 30000 iterations.

wasidennis commented 4 years ago

Please make sure that the pre-trained weight is successfully loaded (it also happened to us that the performance could drop to 2x if not doing so). Normally, the model should have 3x IoU in 10k iterations.

seominseok0429 commented 4 years ago

I'll share our code. (soon)

VE-yyq commented 3 years ago

I'll share our code. (soon)

I have the same question for reproducing. Is there anyone who can share the weights maybe?thanks.