These are images that contain the Dockerfiles for pre-built docker images that contain recent versions of Python + OpenCV + Tensorflow + BigML's Sensenet.
The easiest way to consume these images is to use the uploaded to the official registry.
docker run -it --rm aclementev/python-opencv-tf-sensenet:py3.9-cv4.5.5-tf2.8.0-amd64Which will leave you in a python session.
Note the amd64 at the end of the tag, you must change that to arm64 if you are running a computer with arm CPUs.
The following are the instructions for building the images in a similar way as they are built in the official registry (in Docker Hub).
We use the --platform flag for Docker's BuildKit build backend (default on Docker Desktop version's, must be manually enabled on Linux) for building platform specific version of the images.
Due to the issues described in the M1 and Tensorflow section, at the moment you cannot build the images cross platform.
The version of the image for the linux/amd64 must be built from a x86_64 native host (Rosetta2 does not work yet).
For x86_64:
docker build -t aclementev/python-opencv-tf-sensenet:py3.9-cv4.5.5-tf2.8.0-amd64 --platform linux/amd64 .For arm:
docker build -t aclementev/python-opencv-tf-sensenet:py3.9-cv4.5.5-tf2.8.0-arm64 --platform linux/arm64 .NOTE: The images are manually tagged with the arch version using BuildKit's default build arg TARGETARCH, which will be amd64 on x86_64 and arm64 on arm architectures.
That way you can conditionally select them any Docker image that inherits from these ones, like this:
FROM clementebigml/python-opencv-tf-sensenet:py3.9-cv4.5.5-tf2.8.0-${TARGETARCH}Currently there's a known issue with qemu on macOS not supporting AVX-512 instructions.
These instructions are used by Tensorflow by default as part of the hardware acceleration techniques.
The python wheels uploaded to PyPI have these compiled in by default, and therefore we cannot simply run pip install tensorflow.
One solution would be to use the community-maintained tensorflow-macos project, however that does not play well well with other dependencies
suffering from a similar issue (looking at you, mediapipe).
Therefore we are looking to use an x86 version of the library that runs on macOS's Rosetta (qemu based).
During the installation some native code is running, which makes it so that for now you cannot do a real cross-compilation (neither from mac to linux nor the other way around).