search

NVlabs / DeepInversion

Official PyTorch implementation of Dreaming to Distill: Data-free Knowledge Transfer via DeepInversion (CVPR 2020)
Other
488 stars 79 forks source link

loss about segmentation task #15

Open raninbowlalala opened 3 years ago

raninbowlalala commented 3 years ago

Thanks for your work! I want to generate images from segmentaion task model(like deeplabv3plus), but I use your original loss got bad result. Could you help me with this issue? Thanks a lot!

hongxuyin commented 3 years ago

Other networks like object detector requires augmentations: please check the new WACV'21 paper here. There may be room for segmentation networks too.

raninbowlalala commented 3 years ago

Other networks like object detector requires augmentations: please check the new WACV'21 paper here. There may be room for segmentation networks too.

Thank you very much! I want to try the data augmentation you mentioned- cutout. Could you share with me about what minsize, maxsize and cut number you use? Do you have plan to opensource the code?

hongxuyin commented 2 years ago

Hi @raninbowlalala we have released the code for DI for object detection here: https://github.com/NVlabs/DIODE let me know if it helps.