''voxel_features'' ,output of self.voxelization(features, coords) are zeros

[nicolexie]

hi, I run as follows:
python train.py configs/shapenet/pvcnn/c0p25.py --device 2

and I found that in pvconv.py:
voxel_features, voxel_coords = self.voxelization(features, coords)
the voxel_features elements are alll zeros.

It seems that the _backend does not work, but the output format is right,
and backend is also build:
''Using /tmp/torch_extensions as PyTorch extensions root...
Detected CUDA files, patching ldflags
Emitting ninja build file /tmp/torch_extensions/_pvcnn_backend/build.ninja...
Building extension module _pvcnn_backend...
ninja: no work to do.
Loading extension module _pvcnn_backend...''

[synxlin]

Hi, @nicolexie. I have fixed check API for PyTorch 1.5 and tried running the command. It should work now for PyTorch 1.5.

[nicolexie]

Hi, @nicolexie. I have fixed check API for PyTorch 1.5 and tried running the command. It should work now for PyTorch 1.5.

Thanks for the reply. I use PyTorch 1.3.0 recently. The output format is right, but the values are all zeros, I have not figure out
the reason.

[synxlin]

Hi, @nicolexie.
I would recommend you to delete _pvccn_backend under /tmp/torch_extensions, and tried again.
For Pytorch 1.3, the check API in this file should be AT_CHECK, while TORCH_CHECK for Pytorch 1.5.

[nicolexie]

Hi, @nicolexie.
I would recommend you to delete _pvccn_backend under /tmp/torch_extensions, and tried again.
For Pytorch 1.3, the check API in this file should be AT_CHECK, while TORCH_CHECK for Pytorch 1.5.

I tried, pytorch 1.3 and1.5 both， but I got the same error

[nicolexie]

cuda

Hi, @nicolexie.
I would recommend you to delete _pvccn_backend under /tmp/torch_extensions, and tried again.
For Pytorch 1.3, the check API in this file should be AT_CHECK, while TORCH_CHECK for Pytorch 1.5.

I tried, pytorch 1.3 and1.5 both， but I got the same error

maybe cuda problem

[zx19950617]

Hi, @synxlin @nicolexie ! I met the same problem.
voxel_features, voxel_coords = self.voxelization(features, coords)
voxel_features are zeros.

Confusedly, the program doesn't raise any warning and error.

My environments are as follows:
pytorch 1.3.1
cuda 10.0
How do you solve this problem? Any suggestion will be appreciated a lot!

Best regards

[nicolexie]

Hi, @synxlin @nicolexie ! I met the same problem.
voxel_features, voxel_coords = self.voxelization(features, coords)
voxel_features are zeros.

Confusedly, the program doesn't raise any warning and error.

My environments are as follows:
pytorch 1.3.1
cuda 10.0
How do you solve this problem? Any suggestion will be appreciated a lot!

Best regards

I changed the cuda version, and
My environments are as follows:
pytorch 1.5
cuda 10.0

[Pirazh]

Hi @nicolexie,

So did changing the CUDA version solved your problem entirely? right now I am using CUDA 11.6 and Pytorch 1.13 and I have the exact same issue, without giving any error, the code runs and produces an all zero tensor for the voxel volume despite that point cloud has non-zero features.

I am wondering if you tried the code with a more recent version of CUDA and Pytorch

[HLJT]

你好@nicolexie,

现在更改 CUDA 版本是否解决了您的问题？现在我正在完全使用 CUDA 11.6 和 Pytorch 1.13，我遇到了几十个问题，没有给出任何错误，代码运行并为体素体积生成全零张量，点云具有非零特征。

我想知道您是否使用 CUDA 和 Pytorch 的更新版本尝试过该代码

Hello, I also encountered the same problem as you. Did you solve it? I would be very grateful if you could reply.