Hi, I used your pretrain model directly test linear accuracy on modelnet40, it got 90.27%, same as I runed train_crosspoint.py without any initialize, But the result you mentioned in your paper can get 91.2%. So I want to know are there any tricks in your codes. Or It means I should train based on your pretrain model? I look forward to your answers

Hello, dear author! How did you get the 2D images corresponding to the ModelNet40, ScanObjectNN point cloud data? The content inside eval_ssl.ipynb looks incomprehensible, can you provide the original .py file code?

thanks for your great work! I'm confused that why you fit a simple linear SVM classifier on the train split of the classification datasets in 3D object classification? where can I find the corresponding code?

Hi, thank you for sharing such excellent results

I would like to ask what variant do we use in few-shot learning on ScanObjectNN?

OBJ ONLY OBJ BG PB T25 PB T25 R PB T50 R PB T50 RS

Looking forward for your response, thank you

When using the download_data.sh, it will raise the error: requests.exceptions.MissingSchema: Invalid URL '': No scheme supplied. Perhaps you meant http://?

How to use gdown to download the dataset?

@MohamedAfham Thank you for releasing the code. The paper is well written and the code is robust.

I have successfully trained the classification and part segmentation models based on train_crosspoint.py and train_partseg.py, respectively. Everything goes smoothly.

One point I'm confused with is the comments in scripts/script.sh, you point out train_crosspoint.py can be used for training the part segmentation model and train_partseg.py is used for finetuing it. The code in train_crosspoint.py, however, only load ShapeNetRender for pretraining and ModelNet40 for linear accuracy evaluation. Actually, it does not load ShapeNetPart for part segmentation.

Instead, I think both training and finetuning take place in train_partseg.py as the train_loader in this file is designed for ShapeNetPart. Further, I think the self-superviesd cross-modal contrastive learning is intended for point cloud classification. Have I got a correct understaning?

As the title described, I wonder the GPU device you used to support the batch_size=20.

I use a RTX 2080 Ti, which has 11GB memory, when running train_crosspoint.py, I have to set batch_size=2 to avoid CUDA out of memory since you konw, knn and torch.cat in models/dgcnn.py will consume a large portion of memory.

However, the small batch_size leads to much slower training procedure so that I can get the final results probably in 4 or 5 days.

By the way, I have multiple GPUs, is it possible to incorporate DistributedDataParallel to accelerate the training procedure?

Anyway, I will try it out!

Thank you for your contributions. Very interesting work!. I have two GPUs and I'm trying to run crosspoint pre-training for classification using:

python train_crosspoint.py --model dgcnn --epochs 100 --lr 0.001 --exp_name crosspoint_dgcnn_cls --batch_size 20 --print_freq 200 --k 15

And i'm getting the following error:

Traceback (most recent call last):
  File "train_crosspoint.py", line 258, in <module>
    train(args, io)
  File "train_crosspoint.py", line 100, in train
    _, point_feats, _ = point_model(data)
  File "/home/nas/anaconda3/envs/crosspoint/lib/python3.8/site-packages/torch/nn/modules/module.py", line 1102, in _call_impl
    return forward_call(*input, **kwargs)
  File "/home/nas/Desktop/CrossPoint/models/dgcnn.py", line 95, in forward
    x = get_graph_feature(x, k=self.k)
  File "/home/nas/Desktop/CrossPoint/models/dgcnn.py", line 31, in get_graph_feature
    idx = idx + idx_base
RuntimeError: Expected all tensors to be on the same device, but found at least two devices, cuda:0 and cuda:1! 

Is there a reason for hardcoding cuda device to 1 here: https://github.com/MohamedAfham/CrossPoint/blob/440e3bdf1656014eb4284786a6b2bcdf83e8df30/models/dgcnn.py#L27

Hi, Mohamed Afham! I really appreciate your great work! And I think the figures in your paper are wonderful! Could you please tell me what the pointcloud visualization software is in Figure 2? It's looks nice! Thanks in advance!

Thanks for your nice work. I am trying to reproduce your fine-tuning results on ShapeNetPart segmentation. I find that the model architecture for classification pertaining and segmentation pertaining are different. More specifically, in classification pertaining, the dgcnn model is adopted, while dgcnn_seg is utilized for the pre-training for part segmentation, as shown in the following: https://github.com/MohamedAfham/CrossPoint/blob/364987ed8ae3b9f439b2c08305746ad3c451a820/scripts/script.sh#L6

However, I can not find the definition of dgcnn_seg in your model library. I guess the dgcnn_seg should be the DGCNN_partseg model with pretrain=True, right?

In addition, in my opinion, other paper may adopt the same architecture in pre-training for both classification and part segmentation, such as OcCo. Such a difference may lead to unfair comparison. What's your opinion?

@MohamedAfham Recently, I have run all experiments in the codebase at least 3 times to ensure there are not explicit exceptions during my operations.

Some of the results are very encouraging, which means they are comparable with the paper reported, sometimes even higher than that in the paper, e.g. the reproduced results on ModelNet. But some are not.

Specifically, for the downstream task few-shot classification on ScanObjectNN, the performance gap is relatively large, e.g.,

1. for 5 way, 10 shot, I got 72.5 ± 8.33,
2. for 5 way, 20 shot, I got 82.5 ± 5.06,
3. for 10 way, 10 shot, I got 59.4 ± 3.95,
4. for 10 way, 20 shot, I got 67.8 ± 4.41

For the downstream task linear SVM classification on ScanObjectNN, the reproduced performance is 75.73%. All experiments use the DGCNN backbone and default settings except for the batch size.

In short, all of results are behind the reported peformances on ScanObjectNN in the paper, by a large margin.

At this point, I wonder whether there are some precautions when experimenting on ScanObjectNN, and what possible reasons are. Can you provide some suggestions? thank you.

@MohamedAfham I have succefully integrated the PyTorch DistributedDataParallel mechanism into your codebase, which accelerates the training procedure remarkbly and achieves a similar performance with the paper reported.

Later on I want to pull a request to your repository, thank you.