search

liutinglt / CE2P

214 stars 41 forks source link

Poor result with the pretrained model LIP_epoch_149.pth downloaded from GoogleDrive. #48

Closed maguro27 closed 3 years ago

maguro27 commented 3 years ago

I also encountered the issue as well as https://github.com/liutinglt/CE2P/issues/34. However, this issue does not provide a solution. I do not change the part of weight loading and it does not output any error messages. How do I fix it?

I investigated the output of the intermediate layers. As a result, the output of "self.layer5" goes to -inf in the ResNet class. Moreover, the network generates -inf though I exchange ABN to BN. Thanks.

My environment GPU: Tesla V100 PyTorch: 1.6 ABN: https://github.com/mapillary/inplace_abn Pre-trained model: LIP_epoch_149.pth downloaded from GoogleDrive

CharlesPikachu commented 3 years ago

we re-implement CE2P in this repo: https://github.com/SegmentationBLWX/sssegmentation, which has obtained better performance compared this repo. Maybe you can try this

maguro27 commented 3 years ago

Thank you for your response. It is great work! I try to use your repository. Thanks.