search

THU-DA-6D-Pose-Group / GDR-Net

GDR-Net: Geometry-Guided Direct Regression Network for Monocular 6D Object Pose Estimation. (CVPR 2021)
https://github.com/THU-DA-6D-Pose-Group/GDR-Net
Apache License 2.0
277 stars 47 forks source link

How to choose the best model from the saved checkpoint? #79

Closed zeal-up closed 2 years ago

zeal-up commented 2 years ago

Hi, thanks for releasing your code, that helps me a lot! Have you found the results of the last several saved checkpoints are different and the best result on the test dataset is always not the model_final.pth? Do you make choices from the last several saved checkpoint, or just choose the result of the last model(model_final.pth) to report?

lu-ming-lei commented 2 years ago

hi, what is the specific information of your experiment? In my experiments, I found that epoch needs to be added. According to the epoch given by the author, it often does not reach the optimal convergence.

zeal-up commented 2 years ago

@lu-ming-lei Hi, I follow the parameters provided by the author. I did find that the experiment's result is slightly lower than the paper(although very close). So I think maybe the reported results are not the last model. Do you mean that by increasing the epochs, the result will be better? How much epoch did you add and how much benefit did you get?

lu-ming-lei commented 2 years ago

I also found the same problem, which is slightly lower than the value given in the article. I mean, the results near model_final.pth will fluctuate, such as add (s)

wangg12 commented 2 years ago

Hi, thanks for releasing your code, that helps me a lot! Have you found the results of the last several saved checkpoints are different and the best result on the test dataset is always not the model_final.pth? Do you make choices from the last several saved checkpoint, or just choose the result of the last model(model_final.pth) to report?

We always used the last checkpoint since we do not have a validation set. But you can use some validation set (for example, split some samples from the original training set for validation) to help you choose a best model.