Gengzigang comments

Results 17 comments of


                                            Gengzigang

about offset loss, difficult to convergence?

I have not met this problem. I do not know if you had modified the code and you can try to run the code without any modification.

about offset loss, difficult to convergence?

I used 4GPUs for HRNetw32, and 8GPUs for HRNetw48. The total batch size is 40.

How to make the GT offset maps?

Hi, you can find the construction of GT offset map in the lib/dataset/target_generators/target_generators.py. The center heatmap is to find the center points of each person in the image.

When I was running the training of crowdpose dataset, there were 'nan' and 'NaN or Inf found in input tensor' problems. I checked the data. This problem is caused by the disappearance or explosion of the gradient. After I adjusted the learning rate, this problem still can not be solved. Please help me.

Hi, I am sorry that I have not met this problem. You can reproduce our results in the paper using the cloned code without any change.

Replicate experiments for W48

Hi, you can try our well-trained model in the model zoo. [https://mailustceducn-my.sharepoint.com/personal/aa397601_mail_ustc_edu_cn/_layouts/15/onedrive.aspx?id=%2Fpersonal%2Faa397601%5Fmail%5Fustc%5Fedu%5Fcn%2FDocuments%2FDEKR%2DDisentangled%2DKeypoint%2DRegression%2Fmodel%2Fpose%5Fcoco](url)

Replicate experiments for W48

I cloned the code and tested the model recently, and the result is the same as the paper. You can use the default parameters and the original code to test...

how's the model performance if train on wholebody dataset？

Hi, we did not try to train on wholebody dataset and you can try it.

Why my debug image output is all black?

> I met the same situation,did you know it。 Maybe the learning rate is too high.

How to use Web demo

Thank you for your interest. Due to having only one GPU to support the web demo, it may fail when there are too many people in the queue. You can...

Supported training phase

Thank you for your interest. Regarding pretraining adaptation and human alignment, we have indeed only released the checkpoints. If you would like to train the other two parts, the training...