it would be much more easier (and more friendly) for the newcomers if you push a pre-built Docker images to the Docker Hub: one with GPU and one without. (Or if you manage to auto-build them in there fitting Docker Hub limits)
If those images already exist then ignore this post, but, please, add the information about the Docker Hub image name & tags in your “Install” section on the pyro.ai .
nvidia-container-toolkit makes it a bit easier to build and also to run your images by adding
--gpus argument in
docker run --gpus ...
Personally, on Ubuntu 19.10 (or 20.04) with installed
nvidia-container-toolkit I find it much more easier to start with smth like this:
FROM tensorflow/tensorflow:latest-gpu-py3-jupyter ARG DEBIAN_FRONTEND=noninteractive #RUN apt list --installed RUN apt update && apt install -yq python3-pandas RUN pip3 install pyro-ppl
build docker image with:
docker build -t my-py3-pytorch-pyro-tensorflow2-jupyter .
run docker image:
docker run -u $(id -u):$(id -g) --gpus all -it --rm -v $(realpath ~/path/to/your/notebooks):/tf/notebooks -p 8888:8888 my-py3-pytorch-pyro-tensorflow2-jupyter
(this is especially good for those who also need Tensorflow2 OOTB)