search

Dootmaan / DSRL

Implementation of CVPR 2020 Dual Super-Resolution Learning for Semantic Segmentation
MIT License
95 stars 17 forks source link

I can't seem to trained the model to get good result #17

Open gnsci opened 2 years ago

gnsci commented 2 years ago

Hi there, Thanks for sharing your codes. I wonder if the experience setting in your current train.py can achieve mIoU=0.6768 on Cityscapes. I trained the model, but I can get good result. Can you share the setting for getting mIoU=0.6768?

Dootmaan commented 2 years ago

Hi there, Thanks for sharing your codes. I wonder if the experience setting in your current train.py can achieve mIoU=0.6768 on Cityscapes. I trained the model, but I can get good result. Can you share the setting for getting mIoU=0.6768?

hi @gnsci i think maybe u can try start the training with lr=0.01, then load the weights and restart the training with a lower lr (say 0.001). By the way, the pretrained weights is provided in this repo, u can download it from google drive and directly load it into the network. however, note that this is NOT the official implementation of DSRL so u may find the result is not so good as what is presented in the paper.

gnsci commented 2 years ago

Thanks a lot!