search

plemeri / InSPyReNet

Official PyTorch implementation of Revisiting Image Pyramid Structure for High Resolution Salient Object Detection (ACCV 2022)
MIT License
321 stars 61 forks source link

Is there Res2Net50 pretrained checkpoint for DIS5k dataset? #19

Closed jiamingNo1 closed 1 year ago

jiamingNo1 commented 1 year ago

Hi, I see that the model zoo doesn't have Res2Net50 pretrained checkpoint except for DUTS dataset. Or can you provide relevant training configurations?

plemeri commented 1 year ago

Hello, we did not trained Res2Net50 backbone model for high-resolution datasets. However, you can train the model by changing the configuration Model.name as InSPyReNet_Res2Net50 instead of InSPyReNet_SwinB.

Also, for high-resolution training, change Model.base_size and Train.Dataset.transforms.static_resize.size as [1024, 1024] instead of [384, 384]for larger input size. Don't forget to change datasets by changing Train.Dataset.sets.

Lastly, for high-resolution inference, change Test.Dataset.transforms.static_resize.size as [1024, 1024], or use dynamic_resize with L: 1280.

For more information, please refer to getting started / train & evaluate.

jiamingNo1 commented 1 year ago

Thank you~