MNIST training using nvidia + docker

We use the classical MNIST machine learning problem to bench hardware.

Keras

Keras framework is used in this part.

CPU

To build the CPU image (behind a proxy) :

docker build -t keras-mnist-cpu \
  --build-arg http_proxy=${http_proxy} \
  --build-arg https_proxy=${https_proxy} \
  --build-arg ftp_proxy=${ftp_proxy} \
  --build-arg no_proxy=${no_proxy} \
  keras/cpu

and run it :

docker run \
       -e http_proxy=${http_proxy} \
       -e https_proxy=${https_proxy} \
       keras-mnist-cpu

GPU

First we need to install docker-nvidia, see here. On CentOS 7 :

distribution=$(. /etc/os-release;echo $ID$VERSION_ID)
curl -s -L https://nvidia.github.io/nvidia-docker/$distribution/nvidia-docker.repo | \ 
sudo tee /etc/yum.repos.d/nvidia-docker.repo
sudo yum install -y nvidia-docker2
sudo pkill -SIGHUP dockerd

Capability <= 3.0

We have to compile tensorflow from sources to enable old GPU capabilities :

docker build -t keras-mnist-old-gpu \
  --build-arg http_proxy=${http_proxy} \
  --build-arg https_proxy=${https_proxy} \
  --build-arg ftp_proxy=${ftp_proxy} \
  --build-arg no_proxy=${no_proxy} \
  keras/old-gpu

and run it :

docker run --runtime=nvidia \
       -e http_proxy=${http_proxy} \
       -e https_proxy=${https_proxy} \
       keras-mnist-old-gpu

Capability > 3.0

To build a GPU image (behind a proxy) :

docker build -t keras-mnist-gpu \
  --build-arg http_proxy=${http_proxy} \
  --build-arg https_proxy=${https_proxy} \
  --build-arg ftp_proxy=${ftp_proxy} \
  --build-arg no_proxy=${no_proxy} \
  keras/gpu

and run it :

docker run --runtime=nvidia \
       -e http_proxy=${http_proxy} \
       -e https_proxy=${https_proxy} \
       keras-mnist-gpu

Caffe

Caffe framework is used in this part.

CPU

To build the CPU image (behind a proxy) :

docker build -t caffe-mnist-cpu \
  --build-arg http_proxy=${http_proxy} \
  --build-arg https_proxy=${https_proxy} \
  --build-arg ftp_proxy=${ftp_proxy} \
  --build-arg no_proxy=${no_proxy} \
  caffe/cpu

and run it :

docker run \
       -e http_proxy=${http_proxy} \
       -e https_proxy=${https_proxy} \
       caffe-mnist-cpu

CUDA arch 60 cards only

To build a GPU image (behind a proxy) :

docker build -t caffe-mnist-gpu \
  --build-arg http_proxy=${http_proxy} \
  --build-arg https_proxy=${https_proxy} \
  --build-arg ftp_proxy=${ftp_proxy} \
  --build-arg no_proxy=${no_proxy} \
  caffe/gpu

and run it :

docker run --runtime=nvidia \
       -e http_proxy=${http_proxy} \
       -e https_proxy=${https_proxy} \
       caffe-mnist-gpu