Question about npy file generated by preprocess_data.py

[WangZYZZ]

I try to use preprocess_data.py to generated processed npy file. But the data in npy file makes me confused. [[ 0.2485581 -0.29567248 0.1667352 -0.85212744 0.42836872 0.30063109] [ 0.2485581 -0.29567248 0.1667352 -0.85212744 0.42836872 0.30063109] [ 0.2485581 -0.29567248 0.1667352 -0.85212744 0.42836872 0.30063109] ... [ 0.2485581 -0.29567248 0.1667352 -0.85212744 0.42836872 0.30063109] [ 0.2485581 -0.29567248 0.1667352 -0.85212744 0.42836872 0.30063109] [ 0.2485581 -0.29567248 0.1667352 -0.85212744 0.42836872 0.30063109]]

[2. 2. 2. ... 2. 2. 2.]

I think for each npy file, there should be different coordinates and different labels. But for the npy file above, only label 2 exists and all coordianates are the same. How can i solve it ?

[WangZYZZ]

And I met an error when running command of 'train the Farthest Point Sampling model'

[limhoyeon]

maybe this issue is related to https://github.com/limhoyeon/ToothGroupNetwork/issues/4. there are issues with CUDA functions not executing properly in some PC environments. I plan to remove the CUDA functions and replace them with Python code. Please bear with me a little longer, or alternatively, you can try reinstalling the 'pointops' library from https://github.com/POSTECH-CVLab/point-transformer/tree/master/lib/pointops.

[WangZYZZ]

Before I run 'labeled_vertices = gu.resample_pcd([labeled_vertices], 24000, "fps")[0]' in preprocess_data.py, the labeled_vertices is like The data seems to be ok. But after using 'labeled_vertices = gu.resample_pcd([labeled_vertices], 24000, "fps")[0]', the labeled_vertices is like So is it the problem of gu.resample_pcd ?

[WangZYZZ]

For the pointops, I remeber I have installed it using the setup.py

[WangZYZZ]

should I change anything of the def resample_pcd ?

[WangZYZZ]

Hi, I think i have solved the problem after reinstalling the pointops from your URL. May I know why the pointops can cause this question ?

[limhoyeon]

It seems that I have made some wrong modifications in the PointOps library in this repository,,

[WangZYZZ]

Hi, I have started training the Farthest Point Sampling model from last Friday. But it is still running. May I know how much time did you use to train it ?

[limhoyeon]

Maybe it is hard to track experiments without wandb.

Using wandb, you can track the experiments, as shown below. I trained provided models for 60 epochs(24h in rtx2080).

[WangZYZZ]

Hi, these are my wandb graphs. Is it right? I think there are redundant steps because the lines are almost horizontal after 60 steps. May I know where i can change the number of step or epoch?

[WangZYZZ]

Does the running of this train require me to manually stop it? There is no progress bar on the training screen, and I don't see where to set the STEP number.

[limhoyeon]

I looked at wandb graph and manually stopped the train. you can change the scheduler step(in train_configs/modelname.py) and change the learning rate for each epoch(step_lr graph in wandb).