New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom

Deeplab Training Detail #271

Open

donhk1011 opened this issue Sep 21, 2019 · 0 comments

donhk1011 commented Sep 21, 2019 •

edited

Loading

Hi,

I could not reproduce the deeplab baseline on cityscapes. The baseline is getting ~66% lower than 70.4% in the paper.

I found there is cropping in this repo, https://github.com/msracver/Deformable-ConvNets/blob/master/experiments/deeplab/cfgs/deeplab_resnet_v1_101_cityscapes_segmentation_base.yaml#L60)
But, in the paper, it seems to use the whole image as input in training time.

In addition, the learning rate is different. (5e-4 (https://github.com/msracver/Deformable-ConvNets/blob/master/experiments/deeplab/cfgs/deeplab_resnet_v1_101_cityscapes_segmentation_base.yaml#L52) vs 1e-3 in the paper)

Could you please specify the differences between this repo and the paper?

And is it possible to reproduce the number of the deeplab baseline with the same configuration of this repo?

Thanks!

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment