search

google-research / deeplab2

DeepLab2 is a TensorFlow library for deep labeling, aiming to provide a unified and state-of-the-art TensorFlow codebase for dense pixel labeling tasks.
Apache License 2.0
998 stars 157 forks source link

Adding class weights and training with resnet 18 #101

Closed nithinme3 closed 2 years ago

nithinme3 commented 2 years ago

Hi,

I wanted to add class weights to the existing code according to number of total pixels each classes have, so that the training results will be improved. So where I need to modify the codeto incorporate this? Also I could find pretrained network starting from resnet50 from the documentation provided. Will I be able to train with resnet18 with the existing repository?

aquariusjay commented 2 years ago

Hi @nithinme3,

Please check the losses implemented in base_loss.py. We don't have ResNet-18 ImageNet pretrained checkpoint at hand. If you would like to try small model variants, maybe you could take a look at MobileNetv3-{small,large}.

Cheers,

aquariusjay commented 2 years ago

Closing the issue, since there is no active discussion for a long time. Please feel free to reopen it if needed.