search

lorenmt / reco

The implementation of "Bootstrapping Semantic Segmentation with Regional Contrast" [ICLR 2022].
https://shikun.io/projects/regional-contrast
Other
162 stars 25 forks source link

Random Seed for cityscape #36

Closed ChandayHunt closed 1 year ago

ChandayHunt commented 1 year ago

Hello author, Thank you for your work. I have used the provided code to reproduce the results of cityscape with label 20. I tried the following python command:

python train_semisup.py \
--dataset cityscapes \
--num_labels 20 \
--apply_aug classmix \
--apply_reco \
--seed 0

and I chose --seed from 0, 1 and 2, however the results that I got were around 45% IOU. Could you please share your random seed for inference please? And if you find the command wrong, please correct me.

Thank you!

lorenmt commented 1 year ago

Hi, this looks correct. However, I also found the performance might vary across different machines...might due to different versions of dependencies or so. But the ranking stays the same, the issue will be alleviated with a larger amount of labelled images.

ChandayHunt commented 1 year ago

Yeah that is true. I haven't tried other methods in the table but it's really weird that the IOU didn't reach 49. Anyway thank you for sharing.