yongheng1991 / 3d-point-capsule-networks Goto Github PK

I have the following error after running python build.py install from models/nndistance:

nvcc fatal : Unsupported gpu architecture 'compute_75'

error: command '/usr/bin/nvcc' failed with exit status 1

I have also tried to add the lines below in build.py, but it didn't work:
extra_compile_args={
'cxx': ['-O2',],
'nvcc':['--gpu-architecture=compute_70','--gpu-code=sm_70','-O3','-I./cutlass/','-U__CUDA_NO_HALF_OPERATORS__', '-U__CUDA_NO_HALF_CONVERSIONS__']
}

++++++++++++++++++++++++
Environment:
System: ubuntu 18.04
Python: 3.6.9
PyTorch: 1.3.1
CUDA Version: 10.2
GPU: GeForce RTX2080Ti

There are some problems in running viz_reconstruction.py. The errors are probably due to the recent version of Open3D (0.9.0).

1. Importing open3d after PyTorch causes free(): invalid pointer Aborted (core dumped). So, from open3d import * should be before import torch.

2. This line should be added (before import torch):

import open3d as o3d

Furthermore, these replacements are required in different parts of the code:

PointCloud() ---> o3d.geometry.PointCloud()
Vector3dVector(prc_r_all) ---> o3d.utility.Vector3dVector(prc_r_all)
Vector3dVector(current_patch) ---> o3d.utility.Vector3dVector(current_patch)
draw_geometries([colored_re_pointcloud]) ---> o3d.visualization.draw_geometries([colored_re_pointcloud])

I would like to train this model on custom dataset for 3d point cloud classification. Could you guide me how to prepare the dataset? Thanks.

Sorry if this is obvious, but where are the labels for the dataset, i.e. what objects they correspond to? They don't seem to come with it.

Dear doctor:
I meet a quesion, which is "ImportError: /home/csudxy/anaconda3/envs/env1/lib/python3.6/site-packages/my_lib_cuda-0.0.0-py3.6-linux-x86_64.egg/my_lib_cuda.cpython-36m-x86_64-linux-gnu.so: undefined symbol: _ZN2at19UndefinedTensorImpl10_singletonE",
And my torch.cuda.is_available() is true,So, I don't know where the question is.
My environment: cuda9.2,pytorch1.2.
Good luck!
Liu Yang
2020/9/23

Thanks for sharing your amazing code!
When ran through the viz_reconstruction.py,

File "viz_reconstruction.py", line 61, in
pcd_ = PointCloud()
NameError: name 'PointCloud' is not defined

Any idea how to solve it? Thanks for your time

I also have the following error after running python build.py install from models/torch-nndistance:

/usr/miniconda3/envs/huang_torch/lib/python3.7/site-packages/torch/include/pybind11/pybind11.h:1401:51: required from here
/usr/miniconda3/envs/huang_torch/lib/python3.7/site-packages/torch/include/pybind11/cast.h:2108:44: error: no matching function for call to ‘collect_arguments(pybind11::object&, const pybind11::handle&)’
return detail::collect_arguments(std::forward(args)...).call(derived().ptr());
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
/usr/miniconda3/envs/huang_torch/lib/python3.7/site-packages/torch/include/pybind11/cast.h:2087:1: note: candidate: template<pybind11::return_value_policy policy, class ... Args, class> pybind11::detail::simple_collector pybind11::detail::collect_arguments(Args&& ...)
simple_collector collect_arguments(Args &&...args) {
^~~~~~~~~~~~~~~~~
/usr/miniconda3/envs/huang_torch/lib/python3.7/site-packages/torch/include/pybind11/cast.h:2087:1: note: template argument deduction/substitution failed:
/usr/miniconda3/envs/huang_torch/lib/python3.7/site-packages/torch/include/pybind11/cast.h:2094:1: note: candidate: template<pybind11::return_value_policy policy, class ... Args, class> pybind11::detail::unpacking_collector pybind11::detail::collect_arguments(Args&& ...)
unpacking_collector collect_arguments(Args &&...args) {
^~~~~~~~~~~~~~~~~
/usr/miniconda3/envs/huang_torch/lib/python3.7/site-packages/torch/include/pybind11/cast.h:2094:1: note: template argument deduction/substitution failed:
error: command '/usr/bin/nvcc' failed with exit status 1

Here is the last part of the code.

I would be very grateful if you could give me some help.

Environment:
System: ubuntu 18.04
Python: 3.7.7
PyTorch: 1.2.0
CUDA Version: 9.1
GPU: Quadro GP100

When evaluating capsule auto-encoder on test set, you seem to generate the point cloud based on random grids, not regularly spaced grids. Evaluation with random grids can be hard to compare with other methods due to stochastic property. Could you explain why generation is not deterministic during evaluation?

How does the picture of point cloud generate in your paper? Thanks~

@fedassa @ShunChengWu thanks for open sourcing the wonderfull work , i had few queries
Q1 have you trained the architecture on the available other dataset like semanttic Kitti and 3D dataset
Q2 If not trained can we follow the same training pipeline , if trained can you please share the pre-trained model
Q3 can we use the currently pre-trained model to test on custom dataset which less number of point cloud density

Thanks in advance

Hi thanks for releasing your code :)
I am getting this error, in pointcapsnet_ae.py file. Could you please give some pointers on how to solve this?

u_hat = torch.squeeze(torch.matmul(self.W, x[:, None, :, :, None]), dim=-1)
RuntimeError: cublas runtime error : the GPU program failed to execute at /opt/conda/conda-bld/pytorch_1544174967633/work/aten/src/THC/THCBlas.cu:441

If the version of TensorFlow is 2.0.0, some lines of logger.py should be changed in order to monitor the training process of train_ae.py through TensorBoard.

In logger.py line #15:

self.writer = tf.summary.FileWriter(log_dir)

Should be replaced with:

self.writer = tf.summary.create_file_writer(log_dir)

Also, lines #19 and #20:

summary = tf.Summary(value=[tf.Summary.Value(tag=tag, simple_value=value)])
self.writer.add_summary(summary, step)

Should be replaced with:

with self.writer.as_default():
             tf.summary.scalar(tag, value, step=step)
             self.writer.flush()

In the training process (train_ae.py )of 3D Point Capsule Networks, image_summary and histo_summary are not used, so these two methods can be removed from logger.py.

While running train_ae.py, I encountered an error at line 75 in pointcapsnet_ae.py:

u_hat = torch.squeeze(torch.matmul(self.W, x[:, None, :, :, None]), dim=-1)
Exception has occurred: RuntimeError
cublas runtime error : the GPU program failed to execute at /pytorch/aten/src/THC/THCBlas.cu:441

But if i use a try except to rerun this particular line by:

    try:
        u_hat = torch.squeeze(torch.matmul(self.W, x[:, None, :, :, None]), dim=-1)
    except:
        u_hat = torch.squeeze(torch.matmul(self.W, x[:, None, :, :, None]), dim=-1)

This hacky solution works. Anyone know the reason why?

I am using Ubuntu 18.04, cuda 10.0 and Pytorch 1.0.0

Thank you.

Q1: Did you first train the AE on shapenet_part dataset for point cloud reconstruction，then use the trained encoder to get latent capsules that are used for 3D Object Classification?

Q2: Can the shapenet_part_dataset_ae_200.pth (provided by you）be used to get latent capsules that are used for 3D Object Classification?
Thanks!
@yongheng1991

Thank you for releasing such a wonderful project.
I am so interested in your work, and I want to run the source in Docker
where Error occurred with the following command:

python3 eva_seg.py --model ../../checkpoints/shapenet_part_dataset_ae_200.pth --part_model ../../checkpoints/part_seg_1pecent.p
th --class_choice Airplane

ENVIRONMENT: pytorch 1.0, cuda10.0 cudnn7 OR pytorch 1.1.0 cuda10.0 cudnn7.5
I tracked this error with pdb, and found it occurred when importing open3d. Strangely, it sometimes occurred when importing pytorch-related modules.

Any idea is welcomed!
Thank you in advance

Hello, I only have windows platform at present. Can I run it under win10 platform?

hi, really thanks for your amazing work, but I notice that "eval_seg.py",

parser.add_argument('--class_choice',

if class_choice is assigned "Mug", "Airplane", whatever, then in your dataloader, cls is alwayes "0"
https://github.com/yongheng1991/3D-point-capsule-networks/blob/master/dataloaders/shapenet_part_loader.py#L30
and then it affects "objcats", then "object2setofoid",then "iou_oids", then "part_label".

am I right