A good way to get a taste of Swift for Tensorflow language and tools is to set it up with Jupyter with the fastai Swift notebooks. I wanted a quick setup, which the Mac install experience currently not, so instead I installed the release binaries in a Ubuntu container via Docker. The setup process for this scenario is not well documented, so here it is for you / future me.
What we're about to do is install the S4TF 0.4 release and the fastai v3 Swift notebooks on Ubuntu 18.04. Generally we follow the swift-jupyter docker file, but install cpu-only release versions of the packages.
Below are some of the references I looked at:
https://github.com/tensorflow/swift/blob/master/docs/WhySwiftForTensorFlow.md https://github.com/tensorflow/swift/blob/master/docs/DifferentiableFunctions.md https://github.com/tensorflow/swift/blob/master/docs/PythonInteroperability.md
- https://github.com/google/swift-jupyter
- https://github.com/google/swift-jupyter/blob/master/docker/Dockerfile
- "Currently, the only prebuilt toolchains with LLDB Python3 support are the Swift for TensorFlow Ubuntu 18.04 Nightly Builds."
- https://github.com/tensorflow/swift/blob/master/Installation.md#pre-built-packages
- (note that any "kokoro" references on other pages are out of date, these supercede)
If you're already on the supported Ubuntu 18.04 don't really need Docker.
For Docker Hub you'll need to be logged in:
docker login
Pull the Ubuntu image:
docker pull ubuntu:18.04
Output should have some Pull complete messages ending with something like this:
Status: Downloaded newer image for ubuntu:18.04
docker.io/library/ubuntu:18.04
Verify download:
$ docker images
REPOSITORY TAG IMAGE ID CREATED SIZE
ubuntu 18.04 a2a15febcdf3 22 hours ago 64.2MB
$ docker run --rm ubuntu:18.04 uname -a
Linux f0c696b0c42e 4.9.184-linuxkit #1 SMP Tue Jul 2 22:58:16 UTC 2019 x86_64 x86_64 x86_64 GNU/Linux
Some additional notes on Docker command line options:
- https://docs.docker.com/engine/reference/commandline/create/
- https://docs.docker.com/storage/bind-mounts/
Fully qualified path is necessary to get Mac Docker to plumb the mount all the way through to the host. t and i attach to terminal and run as interactive. priv is needed for among other things debuggers to work (necessary for REPL and Jupyter).
$ docker create -t -i --privileged -v $(pwd)/sharedfiles:/shared -p 8889:8888 ubuntu:18.04 bash
4c5c59fb01a2e1c07edf38624acc5f6b541ad3b9c33420e898f1801ade3a2d03
$ export my_s4tf_container=4c5c59fb01a2
$ docker start $my_s4tf_container
4c5c59fb01a2
$ docker ps
CONTAINER ID IMAGE COMMAND CREATED STATUS PORTS NAMES
4c5c59fb01a2 ubuntu:18.04 "bash" 57 seconds ago Up Less than a second 0.0.0.0:8889->8888/tcp romantic_cartwright
Attach to the running container:
docker attach $my_s4tf_container
To detach: ctrl-p,ctrl-q
Install some the dependencies. Note that we skip graphviz due to its X11 dependency.
$ apt update && apt install -y libvorbis-dev libflac-dev libsndfile-dev cmake build-essential libgflags-dev libgoogle-glog-dev libgtest-dev google-mock zlib1g-dev libeigen3-dev libboost-all-dev libasound2-dev libogg-dev libtool libfftw3-dev libbz2-dev liblzma-dev libgoogle-glog0v5 gcc-6 gfortran-6 g++-6 doxygen libsox-fmt-all parallel exuberant-ctags python-powerline python3-pip curl
Set up Python
apt install python-pip
Install Swift dependencies:
$ apt install -y git cmake ninja-build clang python uuid-dev libicu-dev icu-devtools libbsd-dev libedit-dev libxml2-dev libsqlite3-dev swig libpython-dev libncurses5-dev pkg-config libblocksruntime-dev libcurl4-openssl-dev systemtap-sdt-dev tzdata rsync
Install the latest binary release of S4TF:
$ cd ~
$ curl -O https://storage.googleapis.com/swift-tensorflow-artifacts/releases/v0.4/rc4/swift-tensorflow-RELEASE-0.4-ubuntu18.04.tar.gz
$ mkdir swift
$ tar zxf swift-tensorflow-RELEASE-0.4-ubuntu18.04.tar.gz --directory swift
$ echo 'export PATH=~/swift/bin:$PATH' >> ~/.bashrc
$ source ~/.bashrc
Run Swift interpreter to verify install is ok (ctrl-D to exit):
$ swift
Welcome to Swift version 5.1-dev (LLVM af1f73e9e9, Swift 7d157f346b).
Type :help for assistance.
1>
Set up Swift for Jupyter:
$ pip3 install jupyter matplotlib
$ git clone https://github.com/google/swift-jupyter.git
$ cd swift-jupyter
$ python3 register.py --user --swift-toolchain ~/swift --swift-python-library /usr/lib/x86_64-linux-gnu/libpython3.6m.so --kernel-name "Swift"
You should see some JSON printed to the terminal ending at Registered kernel 'Swift' as 'swift'!
Get fastai course v3 notebooks and launch Jupyter:
$ cd /shared
$ git clone https://github.com/fastai/course-v3.git
$ jupyter notebook --allow-root --ip=0.0.0.0 --port=8888
Point your browser to http://localhost:8889 (the port specified when creating the container) and copy/paste the token=... value ‘ where prompted by the Jupyter login page. You can create a new notebook or look at the fastai notebooks visible here:
- http://localhost:8889/tree/course-v3/nbs/swift An example training an MNIST digit classifier is here:
- http://localhost:8889/notebooks/course-v3/nbs/swift/03_minibatch_training.ipynb