Closed donglao closed 2 years ago
Hi @donglao,
I believe the discussion in issue #7 should be helpful (and possibly issue #19). In short, we only trained a 1x1 conv on top of a dilated ResNet backbone since MoCo does not pretrain a decoder. This gives quite good results.
Thank you for clarification
On Fri, Jan 14, 2022, 6:30 AM Wouter Van Gansbeke @.***> wrote:
Hi @donglao https://github.com/donglao,
I believe the discussion in issue #7 https://github.com/wvangansbeke/Unsupervised-Semantic-Segmentation/issues/7 should be helpful (and possibly issue #19 https://github.com/wvangansbeke/Unsupervised-Semantic-Segmentation/issues/19). In short, we only trained a 1x1 conv on top of a dilated ResNet backbone since MoCo does pretrain a decoder. This gives quite good results.
— Reply to this email directly, view it on GitHub https://github.com/wvangansbeke/Unsupervised-Semantic-Segmentation/issues/23#issuecomment-1013169961, or unsubscribe https://github.com/notifications/unsubscribe-auth/AJ45O2ZFQ2MHG65IVHZRYCDUWAXQDANCNFSM5L2M4LQQ . Triage notifications on the go with GitHub Mobile for iOS https://apps.apple.com/app/apple-store/id1477376905?ct=notification-email&mt=8&pt=524675 or Android https://play.google.com/store/apps/details?id=com.github.android&referrer=utm_campaign%3Dnotification-email%26utm_medium%3Demail%26utm_source%3Dgithub.
You are receiving this because you were mentioned.Message ID: <wvangansbeke/Unsupervised-Semantic-Segmentation/issues/23/1013169961@ github.com>
Greetings,
I have a question regarding linear evaluation. It is my understanding that during evaluation, the model loads weights from both pre-trained encoder and pre-trained Deeplab head (that predicts saliency), then re-initialize and finetunes the last linear layer of Deeplab for semantic segmentation. When comparing with MOCO-V2 pre-trained weights, how do you initialize weights for the Deeplab head?
Thanks, Lao