GPU_Tensorflow_Pytorch_ROS_docker

A customized docker image that supports:

• GPU tensorflow 1.5 (python 2.7 and python 3.5)
• Pytorch 1.0 (python 2.7 and python 3.5)
• ROS Kinetic, Gazebo
• NVIDIA CUDA 9.0, CuDNN 7.0, NVIDIA-driver version 390
• Darknet (C++ object detection)

Installation requirements

• NVIDIA-GPU is available: tested with GTX 1080 Ti, NVS 4200M
• Nvidia-docker is installed Check

nvidia-docker --help

If there is an error, there is a need to install nvidia-docker. Please follow this instruction.

Get the built-in docker

• Go to docker_scripts/
• Build the Docker image (run only 1 time)

cd build_docker_image
bash build.sh

Create a docker container (run only 1 time)

Once docker image is built:

bash create_container.sh

This will create a container named gpu_tf_pytorch_ros. Can check its existance by

nvidia-docker container ls -a

From now on, there is no need to run the above commands.

Run the container

bash run_container.sh 

• If container is already running, we can open another terminal (on host machine) and run

bash attach_container.sh

to log in the same container that we are running.

Manually install CuDNN 7.0 (inside the docker)

Go to NVIDIA download to download cuDNN v7.4.2 (Dec 14, 2018), for CUDA 9.0 (NOTE: Download both libCudnn7_7 and libCudnn7_dev as libCudnn7_dev depends on libCudnn7_7)

Install:

dpkg -i libcudnn7_7.4.2.24-1+cuda9.0_amd64.deb                                                               
dpkg -i libcudnn7-dev_7.4.2.24-1+cuda9.0_amd64.deb                                                           
ldconfig                                                                                                     

DONE!