about first_ins_cent_model Problem

[0ZzYy0]

Hello, I used the challenge dataset and your shared model for prediction. However, there seems to be an issue with the ’self.first_ins_cent_model‘ method in the ’grouping_network_module.py‘ file, which results in all 0 values for whole_cls_1.

[limhoyeon]

Can you give me the visualize result of the sampled points(maybe after farthest point sampling in my inference pipeline code)?

[0ZzYy0]

Thank you for your response,The first image is vertices and the second image is sampled_feats.

[0ZzYy0]

I changed 'new_fps' to 'fps' in 'resample_pcd' method,this is sampled_feats、first_results、result_ins_labels、result_sem_labels.

[limhoyeon]

can you tell me what your graphic cards and Cuda version are? This issue may be related to C++ (cuda kernel) function. So maybe you can see that the farthest point sampling or grouping algorithm in point transformer modules dose not work properly.

[0ZzYy0]

The graphics card is RTX 3090

[limhoyeon]

As I said previously, maybe it is related to the cuda function in the point transformer model(original code -> https://github.com/LiyaoTang/contrastBoundary]. I think that there are some problems when executing the cuda function in Rtx 30XX versions.

I apologize for not being able to assist you immediately, as it might take some time to identify the issues with the CUDA function and modify the code.

On Fri, Jun 16, 2023 at 12:15 PM kou @.***> wrote:

牛逼

— Reply to this email directly, view it on GitHub https://github.com/limhoyeon/ToothGroupNetwork/issues/4#issuecomment-1594029524, or unsubscribe https://github.com/notifications/unsubscribe-auth/AQW6T2XLUTP5H766J3E7LUDXLPFVNANCNFSM6AAAAAAZFW4WEI . You are receiving this because you commented.Message ID: @.***>

[0ZzYy0]

Thank you very much for your help. I will continue to try different methods to solve this problem.

[limhoyeon]

If you download the original pointops library, you might solve your problem. plz refer to https://github.com/limhoyeon/ToothGroupNetwork/issues/7