search

YudeWang / deeplabv3plus-pytorch

Here is a pytorch implementation of deeplabv3+ supporting ResNet(79.155%) and Xception(79.945%). Multi-scale & flip test and COCO dataset interface has been finished.
MIT License
625 stars 117 forks source link

Bad Performance #4

Closed chenyzh28 closed 5 years ago

chenyzh28 commented 5 years ago

I have downloaded the deeplabv3plus_xception_VOC2012_epoch60_all.pth，but the val result is only 0.032. Do you know what happened?

YudeWang commented 5 years ago

I retest the provide model deeplabv3plus_xception_VOC2012_epoch60_all.pth and achieve exactly 81.937% mIoU. Maybe you should check the configuration in config.py again about MODEL_BACKBONE, etc.

YudeWang commented 5 years ago

The repository has been update recently and the pretrained model on COCO is removed due to the change of network structure. I will update once achieve.

I will close the issue for no more discussion. You can also contact me by email yude.wang@outlook.com is necessary.

YourGc commented 5 years ago

I get same issue ! And found when i try net.eval() ,the performance will be bad, but I get good performance with code "with torch.no_gard()" So? Is some issue in BN?