skokec/dau-convnet repository overview

Docker image based on nvidia/cuda that provides python TensorFlow framework with DAU-ConvNet plugin⁠.

The following packages/services are installed:

• CUDA and CuDNN (from nvidia/cuda base docker image)
• Python
• TensorFlow (from pip tensorflow-gpu)
• DAU-ConvNet python plugin

Requirements:

• nvidia-docker v2

⁠Usage

Start docker as:

sudo nvidia-docker run -i -t skokec/dau-convnet:1.0-py3.5-tf1.13.1 /bin/bash

Image tags are composed from [DAU_VERSION]-py[PYTHON_VERSION]-tf[TF_VERSION] with the following versions available:

• DAU_VERSION = 1.0
• PYTHON_VERSION = 3.5, 2.7
• TF_VERSION:
• 1.13.1 (base image nvidia/cuda:10.0-cudnn7-devel-ubuntu16.04 - requires newer nvidia drivers with support for CUDA10 on host)
• 1.12.0, 1.11.0, 1.10.1, 1.10.0, 1.9.0, 1.8.0, 1.7.1, 1.7.0, 1.6.0 (base image nvidia/cuda:9.0-cudnn7-devel-ubuntu16.04)

⁠Building

Images are build using this Dockerfile⁠. You can build your own docker image for specific Python and TensorFlow version using:

docker build -t dau-convnet:1.0-py3.5-tf1.13.1 \
                       --build-arg BASE_CUDA_VERSION=nvidia/cuda:10.0-cudnn7-devel-ubuntu16.04 \
                       --build-arg TF_VER=1.13.1 \
                       --build-arg PY_VER=3.5 \
                       --build-arg PY_VER_MAJOR=3 plugin/tensorflow/docker/ 

Note that TensorFlow is installed from pip so CUDA and CUDNN versions must match with version that is available in PyPI TensorFlow repositories.