search

mhamilton723 / STEGO

Unsupervised Semantic Segmentation by Distilling Feature Correspondences
MIT License
728 stars 150 forks source link

Reproduce Cityscapes results #52

Open vobecant opened 2 years ago

vobecant commented 2 years ago

Dear authors,

thank you very much for your work. Could you please provide a config file that would serve to reproduce your Cityscapes results? The values and settings in the train_config.yml differ from those saved in the provided model checkpoint and as written in the paper. For example, in paper it is written that the output dimension is 70 while the Cityscapes model has 100 dimensions on the output. Also, the hyperparameters saved in the model checkpoint lack some of the variables needed for the full configuration.

Thank you very much in advance.

olvrhhn commented 1 year ago

Dear @mhamilton723, thanks for your great work! I would also be interested in reproducing the numbers on cityscapes. @vobecant did you figure out the config parameters?

Thank you in advance for your help!

pitlover commented 1 year ago

Dear @mhamilton723, I tried to reproduce cityscapes result on vit-base model but failed too.

Can you give me a training log or exact detailed config to reproduce?

Thank you for your help in advance!