How to choose the checkpoint without label information?

mhamilton723 / STEGO

Unsupervised Semantic Segmentation by Distilling Feature Correspondences

MIT License

711 stars 142 forks source link

How to choose the checkpoint without label information? #50

Closed dongzhuoyao closed 1 year ago

dongzhuoyao commented 1 year ago

Hi, I saw it in your code, you choose the checkpoint, by 'test/cluster/mIOU' in pytorch-lightning.

But in your paper, you said that you didn't use any label information in training.

Is there any misunderstanding for me?

Thanks.

mhamilton723 commented 1 year ago

The checkpointing saves both the top mIoU and the latest model. If I recall correctly we just use the model trained longest in results and you should as well. There's often not much difference in practice between the last model and the best model