- Based on nvidia-docker2 image Ubuntu 16.04 / CUDA 9.1 / cuDNN 7
- Adds CMake 3.10.2 to use improved CUDA CMake integration
- Adds miniconda with two environments to build Python 2.7 and 3.6 OpenCV bindings
- Leverages conda to pull Intel MKL headers and shared libraries
- Adds Eigen 3.3.4
- Adds TBB
- Builds OpenCV with all the above (OpenCV cmake generation downloads several other packages like Intel IPP)
- Builds cv2 python module without CUDA and without Intel MKL to make wheel file slightly more portable (many apt-get package still required)
OpenCV is installed in /opt/opencv-3.4.0
Example:
nvidia-docker run --rm -ti mlamarre/docker-cuda-opencv:latest /bin/bash
/# source activate ocvpy3
(ocvpy3) /# python
>>> import cv2
To call Python scripts using cv2 inside follow this example:
/bin/bash -c "source /opt/conda/envs/ocvpy3/bin/activate ocvpy3 && python setup.py install --yes USE_AVX_INSTRUCTIONS"\
This activates the conda environment with the installed cv2.pyd and runs python from that conda env.
The example above runs a setup.py for another project (took this from a docker building dlib). You can build your own environment using pip install [wheel file]. The setuptools wheel files are installed in /usr/local/etc/wheels