Hello, Thank you for sharing this code, it's amazing! Sorry, I have some question about train.py in sem_seg folder, When I run "sh +x train_job.sh" , cmd show this code: "Traceback (most recent call last): File "train.py", line 289, in train() File "train.py", line 238, in train train_one_epoch(sess, ops, train_writer) File "train.py", line 271, in train_one_epoch ops['pointclouds_phs'][1]: current_data[start_idx_1:end_idx_1, :, :], IndexError: list index out of range"

I check train.py parameters, and find a probably reason for GPU use number: parser.add_argument('--num_gpu', type=int, default=1, help='the number of GPUs to use [default: 2]') I just one NVIDIA 1050Ti, so I change default=2 to 1,is that mean I just buy more graphics card to fix this question? THANKS a lot!

Thanks for your awesome code share

I run the train.py code following readme step by step, but when I run python train.py, there is an error:KeyError: "Unable to open object (object 'data' doesn't exist)", here is details:

I solve all the problem of dependency but above error keep showing

I run the pointnet(https://github.com/charlesq34/pointnet) without error, however, I cannot run dgcnn...

please help me, so I can study about dgcnn more

sorry for my poor english skills

@WangYueFt @syb7573330 I could run the code successfully, but the code is running super slow. The speed is about 10 epochs/day. Do you have any idea about this problem or it is the normal speed for this code?

Dear Wang,

I really liked your paper and thanks for sharing your code. I think there is a potential discrepancy between the training and test setup for part segmentation. It would be great if you can please have a look and clarify a few doubts I have.

• In part_seg/test.py, the point cloud is normalized before feeding into the network. While I don't find this being done in part_seg/train_multi_gpu.py. I feel it might hurt performance. Am I missing something here? source: https://github.com/WangYueFt/dgcnn/blob/master/tensorflow/part_seg/test.py#L185

• What is the purpose of the pc_augment_to_point_num? I understand that you remove the extra-points later but won't the network prediction change upon augmenting extra points? source: https://github.com/WangYueFt/dgcnn/blob/master/tensorflow/part_seg/test.py#L185

Looking forward to your response. Best, Ankit

I run the pytorch code with the script python main.py --exp_name=dgcnn_1024 --model=dgcnn --num_points=1024 --k=20 --use_sgd=True And I always get results slightly worse than the reported results in the paper. I used the best test results in the training process. Especially, for average acc (mean class acc), the gap with the reported ones is larger. Are there any special settings or tricks in running the code? Thanks in advance.

Hi,when I run the tensorflow code.I just got the accuracy of 91.2% .I read the paper published in 2018,the result is as sama sa the baseline .I want to the resaon.thanks!

@WangYueFt I find that you compare the result with baseline in the paper. As you mentioned, the baseline is using fixed knn graph rather dynamic graph. So could you help me explain what is the difference between fixed knn graph and dynamic knn graph?

Hello,thank you for your reply,when I try to run code about sem_seg,I meet this problem,and I have one gpu(8gmemory),can you tell me how to solve this problem?looking forward your reply. InternalError (see above for traceback): Blas xGEMM launch failed : a.shape=[1,4096,3], b.shape=[1,3,4096], m=4096, n=4096, k=3 [[Node: tower_0/MatMul = BatchMatMul[T=DT_FLOAT, adj_x=false, adj_y=false, _device="/job:localhost/replica:0/task:0/device:GPU:0"](tower_0/ExpandDims_1, tower_0/transpose)]]

Hi,

I am trying to reproduce your results showing in the paper with your code but I am not able to do it. Would you mind releasing your trained model for shapenet part segmentation task?

Thanks!

Hi, first, sorry for keep asking about your research..

I'm curious about how to calculate forward time(or operation time?) for some models as shown at Table 3 on your paper.

And does that value means computational time for one epoch?

and What effect did you expect by considering 'categorical vector'? I just wonder how you came up with this interesting idea.

Hi, I am impressed by your research and studying. I have a question for visualizing your segmentation outputs. I did some classification deeplearning models, but this is first time for segmentation.

I want to visualize outptus such as Figure6 and Figure 7 on your paper.

Did you use some tools such as MATLAB?

And what should I use for input for visualize?

Sorry for my poor English skills.

Have a nice day

File "C:\Users\ianph\dgcnn\pytorch\main.py", line 225, in train(args, io) File "C:\Users\ianph\dgcnn\pytorch\main.py", line 40, in train train_loader = DataLoader(ModelNet40(partition='train', num_points=args.num_points), num_workers=8, File "C:\Users\ianph\dgcnn\pytorch\data.py", line 66, in init self.data, self.label = load_data(partition) File "C:\Users\ianph\dgcnn\pytorch\data.py", line 45, in load_data all_data = np.concatenate(all_data, axis=0) File "<array_function internals>", line 180, in concatenate

Train 26, loss: 3.676545, train acc: 0.075407, train avg acc: 0.030953 Test 26, loss: 3.640235, test acc: 0.042139, test avg acc: 0.026000 Train 27, loss: 3.671733, train acc: 0.072358, train avg acc: 0.030758 Test 27, loss: 3.637559, test acc: 0.044976, test avg acc: 0.027750 Train 28, loss: 3.675745, train acc: 0.073272, train avg acc: 0.031713 Test 28, loss: 3.636188, test acc: 0.068071, test avg acc: 0.042000 Train 29, loss: 3.691305, train acc: 0.071545, train avg acc: 0.030454

I plugged the DGCNN model into my semantic segmentation framework in which I use other models like PointNet or PointNet++ without problems. At training time everything is fine and I get pretty good accuracies for my Airborne LiDAR data (here I randomly sample 8192 points for each tile so everything is good). However at test time I want to predict all points inside one tile and I get a memory error for a tile with more than 50000 points.

Aborted (core dumped)

I guess the problem is in the pairwise_distance function. This function calculates a adjacency matrix and I think my gpu memory cant handle an array with the shape of 50000 x 50000. I understand that the tf.matmul function is very fast on gpu but I would like to try a workaround which purely calculates the k nearest neighbors without this huge memory overhead. Is there anything like this? I know how to use a KDTree in normal python but I have not found a way yet to use it with tensorflow placeholders...

I know I can work around this problem by using smaller tiles or downsample my point clouds but I really would like to fix this internally...

I list some basic information about my implementation here:

• python 3.6
• tensorflow 1.15
• batch size is already 1 at test time

Thanks in advance for any tips!

From my point of view, since your implementation didn't use the updated node embeddings as input between epochs, it can be seen as a one layer model, right?

So how to add more layers in your model? Have you ever done some experiments about the performance of different layers?

Dear all,

I am using DGCNN to classify LiDAR pointClouds. I have trained the model using ModelNet40 train data(2048 points, 250 epochs) and results are good when I try to classify objects using ModelNet40 test data.

But when I try to classify real data collected by velodyne sensor the prediction is mostly wrong. Please find the attached example. Most of the times I get output as Plant, Guitar or Stairs. I have shifted my objects to center of the coordinate frame and have normalized the values[-1,1]. I have even tried to clean the boundaries.

Can somebody suggest me what I could be doing wrong?