About training details and performance

[zhaoyang10]

Hello @MandyMo ,
Thank you for releasing the code, it is a really great work.
I trained with this code and got reasonable results on 3d datasets, but terrible results on 2d datasets. The losses all seem stable. Then I tried to change hyperparameters, but the results on 2d datasets were far from good. I don't know whether it's normal or not. Really confused.
How is the performance of your pre-trained model?
What's your performance on training set and testing set of 2d datasets?
Can you share the hyperparameter setting or give some instructions?
Thanks a lot! Following are my training details and results.

This is the default parameters setting from this code, just changed the batch_size_(2d, 3d, adv) to fulfill my GPUs. I trained 7M samples for each dataset, you can regard it as about 1M iterations with batch size of 7.
Losses:

Training results from 3d datasets, they seem reasonable.

Training results from 2d datasets, they seem terrible, but the loss for 2d points keeps stable.

Then I amplified the ratio of 2d loss for 10 times, trying to lower 2d loss. But the result didn't goes the way I wanted. The 2d loss went a little bit lower if it divides 10 comparing the former one, it still much higher than 3d keypoint loss. And the results on 2d datasets were not satisfying.
The losses:

 Training results from 3d datasets:

Training results from 2d datasets:

[zhaoyang10]

@willie1997 @hsyntemiz @mehameha998
I read the histories of issues, how about your training results?
Thanks a lot.

[eng100200]

@zhaoyang10 what do u mean by 2d datasets?

[zhaoyang10]

It means datasets with only 2d lables, such as COCO, lsp and so on.

[eng100200]

@zhaoyang10 ok,,,,i think the model is for 2d input images ,, i did not read this work completely,,but, my understanding is this

[eng100200]

@zhaoyang10 do you use wechat?

[zhaoyang10]

You can send me e-mails. It's not convenient to release my wechat directly.

[eng100200]

@zhaoyang10 my email sm_Adnan21@hotmail.com

[eng100200]

@zhaoyang10 added

[RuiboFan]

Hello @MandyMo , Thank you for releasing the code, it is a really great work. I trained with this code and got reasonable results on 3d datasets, but terrible results on 2d datasets. The losses all seem stable. Then I tried to change hyperparameters, but the results on 2d datasets were far from good. I don't know whether it's normal or not. Really confused. How is the performance of your pre-trained model? What's your performance on training set and testing set of 2d datasets? Can you share the hyperparameter setting or give some instructions? Thanks a lot! Following are my training details and results.

This is the default parameters setting from this code, just changed the batch_size_(2d, 3d, adv) to fulfill my GPUs. I trained 7M samples for each dataset, you can regard it as about 1M iterations with batch size of 7.
Losses:

Training results from 3d datasets, they seem reasonable. Training results from 2d datasets, they seem terrible, but the loss for 2d points keeps stable.

Then I amplified the ratio of 2d loss for 10 times, trying to lower 2d loss. But the result didn't goes the way I wanted. The 2d loss went a little bit lower if it divides 10 comparing the former one, it still much higher than 3d keypoint loss. And the results on 2d datasets were not satisfying.
The losses:

 Training results from 3d datasets:

Training results from 2d datasets:

do you have the training datasets?

[Ethan-cpp]

could you help me run the code? I am terrible. Thank you very much!