Error: undefined symbol: _ZNSt15__exception_ptr13exception_ptr10_M_releaseEv about icon

[ExponentialML]

@ftaker887 I have replaced the bvh-distance-queries with PyTorch3D+Kaolin, thus you don't need to install it anymore. Also, the CUDA version is not limited to 11.0.

I just updated the installation scripts, colab can work well for now.

Sorry for the late reply. Thanks for the updates, cheers!

[YuliangXiu]

colab.sh is designed for online colab, don't use it for your workstation. As for local installation, please refer to installation, step by step.

[ExponentialML]

Your installation documentation were the first I followed before trying the collab.sh.
Here are the steps I took after cloning the repository and creating a conda environment, prior to the install:

1. conda install pip

2. pip3 install torch==1.8.2+cu111 torchvision==0.9.2+cu111 torchaudio===0.8.2 -f https://download.pytorch.org/whl/lts/1.8/torch_lts.html

3. conda install -c conda-forge pyembree

4. pip install -r requirements.txt --user

5. git clone https://github.com/facebookresearch/pytorch3d.git cd pytorch3d && pip install -e .

6. Run one of the example scripts.

It works up until it needs to do the 3D pass (bvh) because that version of pytorch is built on cuda 11.1. On step 2, If I use conda to install pytorch, I get the error:

Error: undefined symbol: _ZNSt15__exception_ptr13exception_ptr10_M_releaseEv

It's most likely caused from the difference in C versions of which conda pytorch and pip install pytorch were built with.

Edit: To add, it's preferable to use condas dependencies with Pytorch as we could choose what cuda version to use, and in this case, it would probably solve the issue.

[YuliangXiu]

1. Please refer to #7 to softlink cuda 11.0, and then install pytorch, pytorch3d, and bvh-distance-queries
2. I always use conda for pytorch installation, never use pip to install it, here I use conda install pytorch torchvision torchaudio cudatoolkit=11.1 -c pytorch-lts -c nvidia
3. undefined symbol: _ZNSt15__exception_ptr13exception_ptr10_M_releaseEv such error often means you need to re-compile / re-install the relevant package
4. I update a lighter requirements.txt, and delete lots of unnecessary files, thus you can have another try

[YuliangXiu]

@ftaker887 If you finally successfully set up ICON in your workstation, please share your experience here, which may be helpful for other users. Also let me know, then I will close this issue. Thanks

[ExponentialML]

Hello. I've followed the steps for installing things via conda, and it's still the same issues.

Installing Pytorch-LTS through PIP, pytorch3d local clone, and bvh-distance-queries doesn't lead to the same compilation errors.

I haven't tried soft linking CUDA yet, and will try this later when I'm able to.

[ExponentialML]

Tried all recommended ways of install and no success. The only way this can work in this specific circumstance is upgrading all dependencies to be compatible with with cuda 11.1, and being compatible with conda pytorch3d without locally building.

Soft linking cuda would require installing too many dependencies on my workstation, and would break compatibility with some of the other projects I'm working on.

Unsolved, but closed and moving on as the solution requires too much grunt work.

[YuliangXiu]

@ftaker887 Just try


The advantage of softlink is, in this way, you can change the CUDA version easily. And the reason why I recommend using conda env, is because conda can create a clean and independent environment, with that, any conda install or pip install won't disturb installed packages used by other projects.

I updated the installation.md and requirements.txt yesterday, for now, there is no python setup.py install for bvh.

[YuliangXiu]

@ftaker887 I have replaced the bvh-distance-queries with PyTorch3D+Kaolin, thus you don't need to install it anymore. Also, the CUDA version is not limited to 11.0.

I just updated the installation scripts, colab can work well for now.