Swift-Jupyter

This is a Jupyter Kernel for Swift, intended to make it possible to use Jupyter with the Swift for TensorFlow project.

Installation Instructions

Option 1: Using a Swift for TensorFlow toolchain and Virtualenv

Requirements

Operating system:

Ubuntu 18.04 (64-bit); OR
other operating systems may work, but you will have to build Swift from sources.

Dependencies:

Python 3 (Ubuntu 18.04 package name: python3)
Python 3 Virtualenv (Ubuntu 18.04 package name: python3-venv)

Installation

swift-jupyter requires a Swift toolchain with LLDB Python3 support. Currently, the only prebuilt toolchains with LLDB Python3 support are the Swift for TensorFlow Ubuntu 18.04 Nightly Builds. Alternatively, you can build a toolchain from sources (see the section below for instructions).

Extract the Swift toolchain somewhere.

Create a virtualenv, install the requirements in it, and register the kernel in it:

python3 -m venv venv
. venv/bin/activate
pip install -r requirements.txt
python register.py --sys-prefix --swift-toolchain <path to extracted swift toolchain directory>

Finally, run Jupyter:

. venv/bin/activate
jupyter notebook

You should be able to create Swift notebooks. Installation is done!

Option 2: Using a Swift for TensorFlow toolchain and Conda

Requirements

Operating system:

Ubuntu 18.04 (64-bit); OR
other operating systems may work, but you will have to build Swift from sources.

Installation

1. Get toolchain

Extract the Swift toolchain somewhere.

Important note about CUDA/CUDNN: If you are using a CUDA toolchain, then you should install CUDA and CUDNN on your system without using Conda, because Conda's CUDNN is too old to work with the Swift toolchain's TensorFlow. (As of 2019-04-08, Swift for TensorFlow requires CUDNN 7.5, but Conda only has CUDNN 7.3).

2. Initialize environment

Create a Conda environment and install some packages in it:

conda create -n swift-tensorflow python==3.6
conda activate swift-tensorflow
conda install jupyter numpy matplotlib

3. Register kernel

python register.py --sys-prefix --swift-python-use-conda --use-conda-shared-libs \
  --swift-toolchain <path to extracted swift toolchain directory>

Finally, run Jupyter:

jupyter notebook

You should be able to create Swift notebooks. Installation is done!

Option 3: Using Docker to run Jupyter Notebook in a container

This repository also includes a dockerfile which can be used to run a Jupyter Notebook instance which includes this Swift kernel. To build the container, the following command may be used:

# from inside the directory of this repository
docker build -f docker/Dockerfile -t swift-jupyter .

The resulting container comes with the latest Swift for TensorFlow toolchain installed, along with Jupyter and the Swift kernel contained in this repository.

This container can now be run with the following command:

docker run -p 8888:8888 --cap-add SYS_PTRACE -v /my/host/notebooks:/notebooks swift-jupyter

The functions of these parameters are:

-p 8888:8888 exposes the port on which Jupyter is running to the host.
--cap-add SYS_PTRACE adjusts the privileges with which this container is run, which is required for the Swift REPL.
-v <host path>:/notebooks bind mounts a host directory as a volume where notebooks created in the container will be stored. If this command is omitted, any notebooks created using the container will not be persisted when the container is stopped.

To improve Docker image building, use the new Docker Buildkit system by either setting the DOCKER_BUILDKIT environment variable or configuring the Docker daemon.json. The simplest way is by prepending DOCKER_BUILDKIT=1 to your docker build command:

DOCKER_BUILDKIT=1 docker build -f docker/Dockerfile -t swift-jupyter .

Option 4: Using Docker to run a Swift kernel connected to your local Jupyter Notebook

As of Jupyter Notebook 6.0, you can use --gateway-url= to specify a separate Jupyter Kernel Gateway. (Or use the nb2kg server extension for pre-6.0 versions of Notebook.) This allows running the Swift for Tensorflow Jupyter kernel in a Docker container while running Jupyter Notebook somewhere else, such as your local machine.

First build the basic Swift kernel Docker image (as above), then build the kernel gateway image based on that:

# from inside the directory of this repository
docker build -f docker/Dockerfile -t swift-jupyter .
docker build -f kernel_gateway/Dockerfile -t swift-kg .

Using the new Docker Buildkit system is recommended, as described in the section above.

Then run the kernel gateway:

docker run -p 9999:9999 --cap-add SYS_PTRACE swift-kg

The functions of these parameters are the same as in the section above.

With the gateway running, start Jupyter Notebook in your notebook directory and pass the URL of your kernel gateway:

jupyter notebook --gateway-url 127.0.0.1:9999

(optional) Building toolchain with LLDB Python3 support

Follow the Building Swift for TensorFlow instructions, with some modifications:

Also install the Python 3 development headers. (For Ubuntu 18.04, sudo apt-get install libpython3-dev). The LLDB build will automatically find these and build with Python 3 support.
Instead of running utils/build-script, run utils/build-toolchain-tensorflow, so that you build a whole toolchain that includes LLDB.

This will create a tar file containing the full toolchain. You can now proceed with the installation instructions from the previous section.

Usage Instructions

Rich output with Python

You can call Python libraries using Swift's Python interop to display rich output in your Swift notebooks. (Eventually, we'd like to support Swift libraries that produce rich output too!)

Prerequisites:

You must use a Swift toolchain that has Python interop. As of February 2019, only the Swift for TensorFlow toolchains have Python interop.

After taking care of the prerequisites, run %include "EnableIPythonDisplay.swift" in your Swift notebook. Now you should be able to display rich output! For example:

let np = Python.import("numpy")
let plt = Python.import("matplotlib.pyplot")
IPythonDisplay.shell.enable_matplotlib("inline")

let time = np.arange(0, 10, 0.01)
let amplitude = np.exp(-0.1 * time)
let position = amplitude * np.sin(3 * time)

plt.figure(figsize: [15, 10])

plt.plot(time, position)
plt.plot(time, amplitude)
plt.plot(time, -amplitude)

plt.xlabel("time (s)")
plt.ylabel("position (m)")
plt.title("Oscillations")

plt.show()

Screenshot of running the above two snippets of code in Jupyter

let display = Python.import("IPython.display")
let pd = Python.import("pandas")

display.display(pd.DataFrame.from_records([["col 1": 3, "col 2": 5], ["col 1": 8, "col 2": 2]]))

Screenshot of running the above two snippets of code in Jupyter

Inline plots

You can display images using Swift too.

%install-swiftpm-flags -Xcc -isystem/usr/include/freetype2 -Xswiftc -lfreetype
%install '.package(url: "https://github.com/IBM-Swift/BlueCryptor.git", from: "1.0.28")' Cryptor
%install '.package(url: "https://github.com/KarthikRIyer/swiftplot", .branch("master"))' SwiftPlot AGGRenderer
%include "EnableJupyterDisplay.swift"

Now you should be able to display images! (Currently only PNG format is supported. You also need to provide the image as a base64 String. Eventually we'd like to support other formats as well.)

For example:

import Foundation
import SwiftPlot
import AGGRenderer

func function(_ x: Float) -> Float {
    return 1.0 / x
}

var aggRenderer = AGGRenderer()
var lineGraph = LineGraph()
lineGraph.addFunction(
    function,
    minX: -5.0,
    maxX: 5.0,
    numberOfSamples: 400,
    label: "1/x",
    color: .orange)
lineGraph.plotTitle = "FUNCTION"
lineGraph.drawGraph(renderer: aggRenderer)
display(base64EncodedPNG: aggRenderer.base64Png())

Screenshot of running the above snippet of code in Jupyter

To learn more about displaying plots using SwiftPlot take a look at the documentation here.

%install directives

%install directives let you install SwiftPM packages so that your notebook can import them:

// Specify SwiftPM flags to use during package installation.
%install-swiftpm-flags -c release

// Install the DeckOfPlayingCards package from GitHub.
%install '.package(url: "https://github.com/NSHipster/DeckOfPlayingCards", from: "4.0.0")' DeckOfPlayingCards

// Install the SimplePackage package that's in the kernel's working directory.
%install '.package(path: "$cwd/SimplePackage")' SimplePackage

The first argument to %install is a SwiftPM package dependency specification. The next argument(s) to %install are the products that you want to install from the package.

%install directives currently have some limitations:

You must install all your packages in the first cell that you execute. (It will refuse to install packages, and print out an error message explaining why, if you try to install packages in later cells.)
%install-swiftpm-flags apply to all packages that you are installing; there is no way to specify different flags for different packages.
Packages that use system libraries may require you to manually specify some header search paths. See the %install-extra-include-command section below.

Troubleshooting %installs

If you get "expression failed to parse, unknown error" when you try to import a package that you installed, there is a way to get a more detailed error message.

The cell with the "%install" directives has something like "Working in: /tmp/xyzxyzxyzxyz/swift-install" in its output. There is a binary usr/bin/swift where you extracted the toolchain. Start the binary as follows:

SWIFT_IMPORT_SEARCH_PATH=/tmp/xyzxyzxyzxyz/swift-install/modules <path-to-toolchain>/usr/bin/swift

This gives you an interactive Swift REPL. In the REPL, do:

import Glibc
dlopen("/tmp/xyzxyzxyzxyz/swift-install/package/.build/debug/libjupyterInstalledPackages.so", RTLD_NOW)

import TheModuleThatYouHaveTriedToInstall

This should give you a useful error message. If the error message says that some header files can't be found, see the section below about %install-extra-include-command.

%install-extra-include-command

You can specify extra header files to be put on the header search path. Add a directive %install-extra-include-command, followed by a shell command that prints "-I/path/to/extra/include/files". For example,

// Puts the headers in /usr/include/glib-2.0 on the header search path.
%install-extra-include-command echo -I/usr/include/glib-2.0

// Puts the headers returned by `pkg-config` on the header search path.
%install-extra-include-command pkg-config --cflags-only-I glib-2.0

In principle, swift-jupyter should be able to infer the necessary header search paths without you needing to manually specify them, but this hasn't been implemented yet. See this forum thread for more information.

%include directives

%include directives let you include code from files. To use them, put a line %include "<filename>" in your cell. The kernel will preprocess your cell and replace the %include directive with the contents of the file before sending your cell to the Swift interpreter.

<filename> must be relative to the directory containing swift_kernel.py. We'll probably add more search paths later.

Running tests

Locally

Install swift-jupyter locally using the above installation instructions. Now you can activate the virtualenv and run the tests:

. venv/bin/activate
python test/fast_test.py  # Fast tests, should complete in 1-2 min
python test/all_test_local.py  # Much slower, 10+ min
python test/all_test_local.py SimpleNotebookTests.test_simple_successful  # Invoke specific test method

You might also be interested in manually invoking the notebook tester on specific notebooks. See its --help documentation:

python test/notebook_tester.py --help

In Docker

After building the docker image according to the instructions above,

docker run --cap-add SYS_PTRACE swift-jupyter python3 /swift-jupyter/test/all_test_docker.py