search

Britefury / cutmix-semisup-seg

Semi-supervised semantic segmentation needs strong, varied perturbations
MIT License
165 stars 23 forks source link

Pretrained models #4

Closed wvangansbeke closed 4 years ago

wvangansbeke commented 4 years ago

Hi @Britefury,

Thank you for making the code publicly available.

Can you share the pretrained models for the deeplabv2 architecture (Cutmix, Cutout, VAT and ICT) please? We are having troubles reproducing the results.

Thanks in advance.

Britefury commented 4 years ago

Hi,

I'm afraid I saved the log files, but not the models. Let me see what I can do about producing these.

wvangansbeke commented 4 years ago

Hi @Britefury,

I'm actually training it myself with a different architecture. I made some hyperparameter changes and the results look better now. So it should be fine in the end.

Thanks.

Britefury commented 4 years ago

Hey cool! :)

Which architecture have you gone for?

wvangansbeke commented 4 years ago

Hi @Britefury,

A deeplabv3 with resnet-50 backbone. A resnet101 takes too long to train for me. I'm getting around 50-56% mIoU for 1% of the labels on PASCAL. This is only for 2 different splits, still want to do a third one. Seems plausible based on the numbers in the paper with a deeplabv3+ and resnet101 backbone I guess.

Cheers, Wouter