Pitfalls in Installing Anygrasp • Axi's Blog

A while ago, I wrote this blog post about how to install AnyGrasp and the pitfalls I encountered. Later, I realized this post could also be helpful to many friends overseas, so I translated it into English (with the help of Gemini ↗). Here you can find how to install AnyGrasp, as well as MinkowskiEngine and PointNet2.
For Chinese readers, you can find the original post here and they are the same.
If you have any questions, please feel free to leave a comment or share this post to who you think needs it.

I was recently configuring AnyGrasp and am documenting the problems I encountered here. My environment is Ubuntu 22.04, CUDA 12.1, and cudnn 9.3.0.

Basic Configuration#

First, here is the link to the AnyGrasp GitHub repository: https://github.com/graspnet/anygrasp_sdk ↗. The Installation section provides brief installation steps, but because its dependency, MinkowskiEngine, has not been updated for a long time, some extra steps are required.

First, set up a conda environment:

conda create -n anygrasp python=3.10
conda install openblas-devel -c anaconda
pip install torch 'numpy<1.23' ninja

bash

MinkowskiEngine#

Next, we can start with the first step, which is configuring MinkowskiEngine:

git clone https://github.com/NVIDIA/MinkowskiEngine.git
cd MinkowskiEngine

bash

Based on experience, you need to configure the following environment variables:

export CXX=c++
export CUDA_HOME=/usr/local/cuda-12.1
export MAX_JOBS=2
export SKLEARN_ALLOW_DEPRECATED_SKLEARN_PACKAGE_INSTALL=True

bash

Regarding the last two, MAX_JOBS is for the “CUDA: Out of memory” issue, and SKLEARN_ALLOW_DEPRECATED_SKLEARN_PACKAGE_INSTALL=True is for an outdated scikit-learn issue. Then, if you run the installation command:

python setup.py install --blas_include_dirs=${CONDA_PREFIX}/include --blas=openblas

bash

The first potential error you might encounter has the core issue error: namespace "thrust" has no member "device". This is essentially due to the library being outdated and incompatible with CUDA 12.X.

Following Issue#543 ↗ in the repository, I found a method that worked for me, which involves adding #include to four different files:

src/convolution_kernel.cuh

#include <thrust/execution_policy.h>

cpp

src/coordinate_map_gpu.cu

#include <thrust/unique.h>
#include <thrust/remove.h>

cpp

src/spmm.cu

#include <thrust/execution_policy.h>
#include <thrust/reduce.h>
#include <thrust/sort.h>

cpp

src/3rdparty/concurrent_unordered_map.cuh

#include <thrust/execution_policy.h>

cpp

After that, you might get the error ModuleNotFoundError: No module named 'distutils.msvccompiler'. If so, run pip install "setuptools <65".

Installing again might lead to another error:

Traceback (most recent call last):
  File "/home/gaoning/miniconda3/envs/anygrasp/lib/python3.10/site-packages/torch/utils/cpp_extension.py", line 2105, in _run_ninja_build
    subprocess.run(
  File "/home/gaoning/miniconda3/envs/anygrasp/lib/python3.10/subprocess.py", line 526, in run
    raise CalledProcessError(retcode, process.args,
subprocess.CalledProcessError: Command '['ninja', '-v', '-j', '2']' returned non-zero exit status 1.

txt

You can edit setup.py:

setup(
    name="MinkowskiEngine",
    version=find_version("MinkowskiEngine", "__init__.py"),
    install_requires=["torch", "numpy"],
    packages=["MinkowskiEngine", "MinkowskiEngine.utils", "MinkowskiEngine.modules"],
    package_dir={"MinkowskiEngine": "./MinkowskiEngine"},
    ext_modules=ext_modules,
    include_dirs=[str(SRC_PATH), str(SRC_PATH / "3rdparty"), *include_dirs],
    cmdclass={"build_ext": BuildExtension.with_options(use_ninja=False)},
    author="Christopher Choy",
    author_email="[email protected]",
    ...,
)

python

Set use_ninja to False, and then run the installation again. It should work without issues.

Note that when installing with CUDA 12.4, an additional error occurred: error: no instance of overloaded function "std::__shared_ptr<_Tp>::_M_enable_shared_from this ...". This triggers several errors, but they all revolve around __shared_ptr and __to_address, specifically related to overloading.

This issue requires modifying /usr/include/c++/12/bits/shared_ptr_base.h (this was the file for me; someone in the Issue mentioned a different one, ptr_traits.h). Search for and replace the following:

auto __raw = __to_address(__r.get());
auto __raw = std::__to_address(__r.get());

cpp

Then, another error might appear: ld: cannot find -lopenblas: No such file or directory; collect2: error: ld returned 1 exit status. After installing conda install openblas-devel -c anaconda, you need to copy a file. For me, the command was cp /ssd/gaoning/miniconda3/envs/anygrasp/lib/libopenblas.so* /ssd/gaoning/miniconda3/envs/anygrasp/lib/python3.10/site-packages/torch/lib/..

After that, the installation should proceed as usual.

AnyGrasp#

Next, install anygrasp_sdk:

git clone https://github.com/graspnet/anygrasp_sdk.git
cd anygrasp_sdk
pip install -r requirements.txt

bash

PointNet2#

Next, install pointnet2:

cd pointnet2
python setup.py install

bash

It’s worth mentioning that during the installation of pointnet2, the Command '['ninja', '-v', '-j', '2']' error might still occur. The solution is the same as before: modify the arguments of the setup() function in setup.py.

During this process, a rather rare issue might also appear:

gcc: fatal error: cannot execute ‘cc1plus’: execvp: No such file or directory
compilation terminated.
error: command '/usr/bin/gcc' failed with exit code 1

txt

Usually, running sudo apt install build-essential is enough, but my problem went beyond that, as gcc and g++ on my system were fine. After checking, I found it’s because the compilation process of pointnet2 involves gcc calling g++. gcc likely calls the same version of g++, and the problem might have been caused by a version mismatch between gcc --version and g++ --version. Reinstall using a specific version with sudo apt install (for Ubuntu 22.04, this can be version 12):

sudo apt install gcc-12 g++-12

bash

After this, the installation should proceed normally.

License Checker#

Finally, using AnyGrasp requires a key, which needs to be generated using ./license_checker -f. On Ubuntu 22.04, this can also cause errors: one is the missing libcrypto.so.1.1, and the other is sh: 1: ifconfig: not found.

# to solve libcrypto.so.1.1 issue
find / -name libcrypto.so.1.1
# for example found a libcrypto.so.1.1 from cuda
sudo ln -s /usr/local/cuda-12.1/nsight-systems-2023.1.2/host-linux-x64/libcrypto.so.1.1 /usr/lib/libcrypto.so.1.1
# to solve ifconfig issue
sudo apt install net-tools

bash

Also, the machine code output after running ./license_checker -f will likely end with a %. This is because a newline character was not added to the output (Python’s print() adds one by default, but the language used for this program might not). Just submit the content without the % to the application form. Additionally, you need to use an educational email address.