Julia & graphic enthusiast. Interested in high performance computing, GPUs and machine learning.

Julia

pkg"up; add WebSockets WebIO#master JSExpr@0.3.1 Observables CSSUtil ImageFiltering Flux@0.6.8 ImageTransformations CuArrays#master GPUArrays#master NNlib#master Makie#master AbstractPlotting#master FixedPointNumbers"
pkg"precompile"
nothing
45.8s
using Flux, Flux.Data.MNIST, Statistics
using Flux: onehotbatch, onecold, crossentropy, throttle
using Base.Iterators: repeated, partition
using CuArrays

# Classify MNIST digits with a convolutional network
imgs = MNIST.images()
labels = onehotbatch(MNIST.labels(), 0:9)

# Partition into batches of size 1,000
train = gpu.([(cat(float.(imgs[i])..., dims = 4), labels[:,i])
         for i in partition(1:60_000, 1000)])

# Prepare test set (first 1,000 images)
tX = gpu(cat(float.(MNIST.images(:test)[1:1000])..., dims = 4))
tY = gpu(onehotbatch(MNIST.labels(:test)[1:1000], 0:9))

reshape4(x) = reshape(x, :, size(x, 4))
maxpool2x2(x) = maxpool(x, (2,2))

m = gpu(Chain(
  Conv((2,2), 1=>16, relu),
  maxpool2x2,
  Conv((2,2), 16=>8, relu),
  maxpool2x2,
  reshape4,
  Dense(288, 10), 
  softmax
))

loss(x, y) = crossentropy(m(x), y)
accuracy(x, y) = mean(onecold(m(x)) .== onecold(y))

evalcb = throttle(() -> @show(accuracy(tX, tY)), 10)
opt = ADAM(Flux.params(m));
# train
for i = 1:10
    Flux.train!(loss, train, opt, cb = evalcb)
end
using ForwardDiff

f(x) = x^2 + 2.0x + 3.0
f_grad(x) = 2x + 2 # manual gradient function

val = 5
dual = ForwardDiff.Dual(val, 1)
dx = f(dual)
@show ForwardDiff.value(dx) == f(val)
@show ForwardDiff.partials(dx)[] == f_grad(val)
using BenchmarkTools
vals = fill(val, 10^6)
duals = ForwardDiff.Dual.(vals, 1)
@btime f($val)
@btime f($dual)

@btime f.($vals)
@btime f.($duals)
nothing
function d3tod2(input3d)
  rotr90(hcat((input3d[ :, :, i] for i in 1:size(input3d, 3))...))
end
function squeeze(A)
  dims = ((i for i in 1:ndims(A) if size(A, i) == 1)..., )
  dropdims(A, dims = dims)
end

function visualize_layer(scene, layer::Union{Conv, typeof(maxpool2x2)}, input)
  img = lift(d3tod2  collect, input)
  heatmap!(scene, 0..16, 0..1, img, colormap = [:black, :white], interpolate = true)
end

function visualize_layer(scene, layer::Dense, input)
  heatmap!(scene, 0..16, 0..1, lift(x-> collect(layer.W'), input))
  height = lift(input) do x
    (collect(x)[:, 1] .+ 15) ./ 15
  end
  xrange = range(1, stop = 15, length = 10)
  barplot!(scene, xrange, height, color = height, colormap = Reverse(:Spectral))
  annotations!(
    scene,
    string.(0:9), Point2f0.(xrange, 0.1),
    textsize = 0.5
  )
end

function visualize_layer(scene, layer, input)
  # ignore unknown layers
end
using Makie, Observables, WebSockets, ImageShow, WebIO
using AbstractPlotting: hbox!
function visualize(m, input)
  scene = Scene(camera = cam2d!, raw = true, resolution = (650, 250))
  input_img = lift(rotr90  squeeze  collect, input)
  image!(scene, (-8)..(-1), (0.0)..7, input_img, colorrange = (0, 1))
  drawto!(scene, input, scene[end])
  output = async_latest(input)
  for layer in m
    output = lift(x-> layer(x), output)
    visualize_layer(scene, layer, output)
  end
  hbox!(scene.plots[3:end])
  center!(scene)
  display(scene)
  scene
end
input_node = Makie.Node(tX[:, :, 1:1, 1:1])
scene = visualize(m, input_node)
lift(x-> (yield();scene), input_node)
evalcb = throttle(0.01) do
  @show(accuracy(tX, tY))
  input_node[] = input_node[] # update node
end
for i = 1:10
    Flux.train!(loss, train, opt, cb = evalcb)
end
index.html
HTML(read(index.html, String))