@linkinpark213 figured out how to install DensePose on CentOS 7 with Anaconda in this brilliant blog post. He is the Sherlock Holmes. @Johnqczhang wrote this brilliant tutorial. This repository follows their instructions. These shell scripts help speed up the installation process. Especially when you fail, it saves time if you start over.

• wants to install DensePose (obviously)
• your computer/server/cluster uses CentOS 7 operation system
• don't want to compile Caffe2 yourself and would like to install Caffe2 from Anaconda

You will fail if you follow the official installation guide. The problems are that:

• Although it is stated in the official Caffe2 installation guide that simply running conda install pytorch-nightly -c pytorch "will install Caffe2 and all of its required dependencies into the current conda environment", it won't.
• DensePose uses some customized operations that need to be manually compiled. You need to compile them with gcc-4.9.2 and protobuf-3.5.0, otherwise they won't work. This is because the Caffe2 you installed from Anaconda (together with pytorch 1.0.0) was originally built with gcc-4.9.2 and protobuf-3.5.0.
• Sadly, gcc-4.9.2 and protobuf-3.5.0 are not available from Anaconda.

My environment:

• CentOS 7.2 cluster
• GTX 1080ti GPU
• Nvidia driver 384.98
• CUDA 9.0
• CuDNN 7.3.1 (installed from Anaconda)
• Anaconda 4.6.8
• Python 2.7.15 (installed from Anaconda)

clone this repository

https://github.com/yzhq97/densepose_centos_installation.git

In 0_set_variables.sh, you need to specify the anaconda environment name and where you would like to install DensePose. It's okay to use default values. It installs DensePose under your $HOME/densepose by default.

DENSEPOSE_DIR="$HOME/densepose" # The place where you want to install DensePose
CONDA_ENV_NAME="densepose" # The name of your anaconda environment
CUDA_DIR="/mnt/lustre/share/cuda-9.0" # path to your cuda installation directory

In CMakeList.txt, you need to specify the path to your Anaconda environment.

set(CONDA_ENV_DIR "/mnt/lustre/yangzhuoqian/anaconda3/envs/densepose")

You can just run sh 1_create_conda_environment.sh. It will create an environment with Python 2.7 and add the CUDA paths into the environment variables of this anaconda environment.

after this, activate the environment: conda activate densepose (if the name of your environment is densepose)

sh 2_download_resources.sh
sh 3_install_dependencies.sh
sh 4_build_gcc492.sh

These scripts could take about an hour, or more.

conda deactivate densepose
conda activate densepose

This is because 4_build_gcc492.sh adds some more environment variables. You need to reactivate the environment for those changes to take effect.

sh 5_build_protobuf.sh
sh 6_build_cocoapi.sh
sh 7_install_pytorch.sh

sh 8_test_caffe2_import.sh

Expected output:

Success
4 # depends on the number of gpus in your system

sh 9_build_detecron.sh
sh 10_test_detectron.sh

Expected test output:

[E init_intrinsics_check.cc:43] CPU feature avx is present on your machine, but the Caffe2 binary is not compiled with it. It means you may not get the full speed of your CPU.
[E init_intrinsics_check.cc:43] CPU feature avx2 is present on your machine, but the Caffe2 binary is not compiled with it. It means you may not get the full speed of your CPU.
[E init_intrinsics_check.cc:43] CPU feature fma is present on your machine, but the Caffe2 binary is not compiled with it. It means you may not get the full speed of your CPU.
...
----------------------------------------------------------------------
Ran 3 tests in 8.237s

OK

sh 11_build_ops.sh
sh 12_test_ops.sh

This is where most people are stuck. Note that the 11_build_ops.sh moves the modified CMakeList.txt into your DensePose directory, replacing the old one. If something reports wrong about CMake, you should probably modify your own CMakeList.txt from this CMakeLists.txt

If everything turns out right, you should see something like:

[E init_intrinsics_check.cc:43] CPU feature avx is present on your machine, but the Caffe2 binary is not compiled with it. It means you may not get the full speed of your CPU.
[E init_intrinsics_check.cc:43] CPU feature avx2 is present on your machine, but the Caffe2 binary is not compiled with it. It means you may not get the full speed of your CPU.
[E init_intrinsics_check.cc:43] CPU feature fma is present on your machine, but the Caffe2 binary is not compiled with it. It means you may not get the full speed of your CPU.
............
----------------------------------------------------------------------
Ran 12 tests in 38.621s

OK

Phew, for a minute there, I lost myself.

• Official DensePose repository
• [MineSweeping] The Long Struggle of DensePose Installation by @linkinpar213, the one who figured it all out.
• The real installation guide from @Johnqczhang