search

pleaseconnectwifi / PNAL

Learning with Noisy Labels for Robust Point Cloud Segmentation (ICCV2021 Oral)
61 stars 6 forks source link

Dataloader of SCANNET is not provided. #5

Closed liujiaheng closed 3 years ago

liujiaheng commented 3 years ago

image

Hi, in "PANL_dgcnn.py", the dataloader of scannet is not provided.

liujiaheng commented 3 years ago

image image util.io_utils'' andmake_NL_S3DIS.py'' are also not provided.

Thanks for your nice work and code.

liujiaheng commented 3 years ago

image Hi, I have another question, the training process of the ``PNAL'' is very slow, is it reasonable?

pleaseconnectwifi commented 3 years ago

Hi Jiaheng,

Glad that you find this work helpful!

As for the code release, I am still working on preparation of it. However it may take some time as I am rushing to deadline of CVPR 2022... 

As for the training speed, in our experiments we trained on 11G 1080Ti or 11G 2080Ti GPUs. For the warm-up stage it takes about 5 mins for 100 iters, and for the next stage it takes less than 15 mins for 100 iters. You can refer to the attached log. According the performace and speed of these GPUs V100 will be faster and much more powerful than 2080Ti. So you might check the envirment and the CPU?

Best,

Shuquan

Pixie8888 commented 2 years ago

Hi @liujiaheng, I encountered same problem of training very slow. The pre-training stage is normal speed but it takes around 30 minutes to train 10 iters on a 1080ti in the cleaning stage. Did you managed to solve this problem?

It seems that there's large portion of time that it operates in the cpu in the cleaning stage and this is very slow.

liujiaheng commented 2 years ago

No