An Introduction for Pythonic Data Scientists

Chad Scherrer

Senior Data Scientist, Metis

Follow Along Live:

https://bit.ly/2FqdelB

What is Julia?

Language Features

Built for performance from day 1
JIT compilation
Easy to call Python, R, C, Fortran
Macros for automating code rewrites
Multiple dispatch

Benefits

Lowers user/dev barrier
Less need to switch languages
Better for pedagogy

People have assumed that we need both fast and slow languages. I happen to believe that we don't need slow languages.

- Jeff Bezanson, Julia co-creator

Plots.jl Ecosystem

Backends

PyPlot

Plotly / PlotlyJS

UnicodePlots

PGFPlots

InspectDR

HDF5

StatsPlots

PlotRecipes

AtariAlgos

Reinforce

JuliaML

Augmentor

DifferentialEquations

PhyloTrees

EEG

ImplicitEquations

ControlSystems

ValueHistories

ApproxFun

AverageShiftedHistograms

MLPlots

LazySets

A Simple Plot

using Plots
using StatsPlots, Distributions
plot(Normal(1,2), legend=false, lw=2

Easy GIFs!

plt = plot3d(1
  , xlim=(-25,25)
  , ylim=(-25,25)
  , zlim=(0,50)
  , title = "Lorenz Attractor"
  , legend=false)

@gif for i=1:3000
    step!(attractor)
    push!(plt
      , attractor.x
      , attractor.y
      , attractor.z)
end every 20

Side Quest

Using the Julia REPL

Measurements.jl

function mySqrt(x)
    ans = (x + 1) / 2
    for _ in 1:10
        ans = (ans + x / ans) / 2
    end
    ans
end
    
x = (0.1 : 0.3 : 4) .± 0.2
y = mySqrt.(x)

scatter(x, y
  , legend = false
  , xlabel = L"x"
  , ylabel = L"\sqrt x"
)

ScikitLearn.jl

using GaussianMixtures: GMM
using ScikitLearn

gmm = fit!(GMM(n_components=3, kind=:diag), X)
predict_proba(gmm, X)

Interface to scikit-learn using PyCall.jl
Easy access to Python methods
All of scikit-learn, plus
- GaussianMixtures.jl
- DecisionTree.jl
- LowRankModels.jl
- Easy to add more!

TensorFlow.jl

i = tf.constant(0, name="i")
result = tf.constant(0, name="result")

output = tf.while_loop(
  lambda i, result: tf.less(i, 10), 
  lambda i, result: [i+1, result+tf.pow(i,2)], 
  [i, result]
)

@tf i = constant(0)
@tf result = constant(0)
output = @tf while i < 10
  i_sq = i^2
  [i=>i+1, result=>result+i_sq]
end

Python

Julia

Lots of advantages over the Python API
Ready for TensorFlow 2.0

https://github.com/malmaud/TensorFlow.jl

MXNet.jl

using MXNet

mlp = @mx.chain mx.Variable(:data)             =>
  mx.FullyConnected(name=:fc1, num_hidden=128) =>
  mx.Activation(name=:relu1, act_type=:relu)   =>
  mx.FullyConnected(name=:fc2, num_hidden=64)  =>
  mx.Activation(name=:relu2, act_type=:relu)   =>
  mx.FullyConnected(name=:fc3, num_hidden=10)  =>
  mx.SoftmaxOutput(name=:softmax)

# data provider
batch_size = 100
include(Pkg.dir("MXNet", "examples", "mnist", "mnist-data.jl"))
train_provider, eval_provider = get_mnist_providers(batch_size)

# setup model
model = mx.FeedForward(mlp, context=mx.cpu())

# optimization algorithm
# where η is learning rate and μ is momentum
optimizer = mx.SGD(η=0.1, μ=0.9)

# fit parameters
mx.fit(model, optimizer, train_provider, n_epoch=20, eval_data=eval_provider)

Flux.jl

https://github.com/FluxML/Flux.jl

encoder = Dense(28^2, N, leakyrelu) |> gpu
decoder = Dense(N, 28^2, leakyrelu) |> gpu

m = Chain(encoder, decoder)

loss(x) = mse(m(x), x)

evalcb = throttle(() -> @show(loss(data[1])), 5)
opt = ADAM()

@epochs 10 Flux.train!(
      loss
    , params(m)
    , zip(data)
    , opt
    , cb = evalcb
)

m = Chain(
  LSTM(N, 128),
  LSTM(128, 128),
  Dense(128, N),
  softmax)

m = gpu(m)

function loss(xs, ys)
  l = sum(crossentropy.(m.(gpu.(xs)), gpu.(ys)))
  Flux.truncate!(m)
  return l
end

opt = ADAM(0.01)
tx, ty = (Xs[5], Ys[5])
evalcb = () -> @show loss(tx, ty)

Flux.train!(loss, params(m), zip(Xs, Ys), opt,
cb = throttle(evalcb, 30))

Autoencoder

Character RNN