你好，请问训练和测试都出现如下错误，是否环境配置原因？

EMD-CD: 0%| | 0/19 [00:00<?, ?it/s] Traceback (most recent call last): File "train_ae.py", line 203, in cd_loss = validate_loss(it) File "train_ae.py", line 169, in validate_loss metrics = EMD_CD(all_recons, all_refs, batch_size=args.val_batch_size, accelerated_cd=True) File "/userhome/point_cloud/diffusion-point-cloud-main/evaluation/evaluation_metrics.py", line 58, in EMD_CD cd_lst.append(dl.mean(dim=1) + dr.mean(dim=1)) RuntimeError: CUDA error: invalid device function Segmentation fault (core dumped)

问题1：

问题2： Exception ignored in: <function _MultiProcessingDataLoaderIter.del at 0x7fb8cf914320> Traceback (most recent call last): File "/root/miniconda3/envs/dpm-pc-gen/lib/python3.7/site-packages/torch/utils/data/dataloader.py", line 1101, in del self._shutdown_workers() File "/root/miniconda3/envs/dpm-pc-gen/lib/python3.7/site-packages/torch/utils/data/dataloader.py", line 1075, in _shutdown_workers w.join(timeout=_utils.MP_STATUS_CHECK_INTERVAL) File "/root/miniconda3/envs/dpm-pc-gen/lib/python3.7/multiprocessing/process.py", line 138, in join assert self._parent_pid == os.getpid(), 'can only join a child process' AssertionError: can only join a child process

Hi folks,

when training the model with "Car" shapes, the process is halted because of the following error

Pairwise EMD-CD: 9%|########6 | 36/400 [00:00<00:03, 97.56it/s] Traceback (most recent call last): File "train_gen.py", line 218, in test(it) File "train_gen.py", line 185, in test results = compute_all_metrics(gen_pcs.to(args.device), ref_pcs.to(args.device), args.val_batch_size, accelerated_cd=True) File "/pointcloudsgan/code/diffusion-point-cloud/evaluation/evaluation_metrics.py", line 196, in compute_all_metrics ref_pcs, sample_pcs, batch_size, accelerated_cd=accelerated_cd) File "/pointcloudsgan/code/diffusion-point-cloud/evaluation/evaluation_metrics.py", line 110, in pairwise_EMD_CD emd_batch = emd_approx(sample_batch_exp, ref_batch) File "/pointcloudsgan/code/diffusion-point-cloud/evaluation/evaluation_metrics.py", line 21, in emd_approx emd = match_cost(sample, ref) # (B,) File "/pointcloudsgan/code/diffusion-point-cloud/evaluation/StructuralLosses/match_cost.py", line 22, in forward cost = MatchCost(seta, setb, match) RuntimeError: CUDA kernel failed : 700

It is quite strange since we use the pretrained model to test airplane/chair generation and don't have any problems using Pairwise EMD-CD. Any insights?

Our GPU is RTX 6000 with CUDA 11.0. StructuralLoss was also correctly compiled.

Many thanks in advance.

How to draw the image generated by the beautiful 3D Point Cloud Generation in this article, can anybody share the code of the drawing? Since i like it very much.

Hi！ I am interested in your wonderful work!

I have a question about training. In your training file "train_gen.py", you set the training iteration number to 'inf' and the training loss is around 60.5, also the grad is around 0.2 to 0.3. So I want to know when to stop straining and I cannot judge stop time just by the loss and grad. Also I want to know why the training loss almost ever decrease a lot.

Looking forward to your response!

Best.

Hi,

I have built "StructuralLoss" successfully. But for the evaluation_metrics.py, I found the problem that it can not import " from evaluation.StructuralLosses.StructuralLossesBackend import ApproxMatch, MatchCost, MatchCostGrad" in match_cost.py ( folder: pytorch_structural_losses) because "StructuralLossesBackend" is not created yet.

So, I tried to go to "pytorch_structural_losses" folder and run the "setup.py" file using the following command:

python setup.py build_ext

But getting errors. ( I have attached a screenshot) . I tried both python 3.7 and 3.8 but getting the same errors.

I tried multiple times but could not able to find it. Please let me know how to fix this issue?

Hi, first thanks for your inspiring work! Point cloud rendering figures in your paper are so beautiful, such as these vivid shadows. How do you draw it? Using open3d, meshlab or other programmes?

Thanks!

Hello:) I got a point cloud, but I don't know how to wrap it in a mesh, I tried a lot of methods, but they all turn out too not pretty (example in the picture). Do you know how to improve the point cloud to get a better mesh shape? or maybe you know cool algorithms for getting a beautiful mesh shape?

Hi, I'm confused about the formula below: I can understand the calculation of entropy by the formula below, However, I cannot understand why the first item is calculated by standard_normal_logprob function, the function looks like a sum of lnf(x). how this function can correspond to the first item?

Thank you

Dear Sir: Thanks for your sharing your elegant work. I am still confused about how to train this model on ShapeNet, even after reading your paper. Could you provide your training code for my better understanding

Guys , I implemented your awesome work to HuggingFace as a demo . People can try your work . You can reach me for any edit

https://huggingface.co/spaces/SerdarHelli/diffusion-point-cloud

Hi! Thanks for your wonderful work and I am following your work.

I find that diffusion model used in image areas usually follow settings in Ho et al.(2020) setting num steps to 1000 and beta T to 0.01/0.02 but in your paper, num steps fixed to 200 and beta T set to 0.05. I just want to know how to finally get those parameters (or just try all kinds of cases) since I can not get reasonable result after changing those parameters some times when I apply your work to other point cloud problem.

Looking forward to your reply! Thanks a lot.

Hi Zhitong, thanks for releasing the code! I am wondering how many training iterations is needed for different model/data to reproduce the results in the paper?

Best,