vankesteren

# Sparse coding for MNIST feature extraction using autodiff
using Zygote: gradient
using MLDatasets: MNIST
using LinearAlgebra: Diagonal
using ImageCore

# Goodfellow page 629, equation 19.16, but per pixel
function loss(H::Matrix{Float64}, W::Matrix{Float64})
    (sum(abs.(H)) + sum((X - H*W).^2)) / L
end

vankesteren

Department compute server

This is the documentation for using the Methods & Statistics department compute server.

• Department compute server
  • Connecting to the server
  • User account
    • Creating a user
    • Updating the password
  • Using the server
• Server specifications

vankesteren

mkdir eps; for file in $(ls | grep .pdf); do pdf2ps $file eps/${file%.*}.eps; done

vankesteren

# Covariance matrix and random normal data
Sigma <- matrix(c(1, .5, .5, 1), 2)
X     <- matrix(rnorm(1000), ncol = 2)
S     <- cov(X)

# random multivariate normal sample
R_mvn <- chol(Sigma)
X_mvn <- X %*% R_mvn
cov(X_mvn)

vankesteren

// cheat on guessthecorrelation.com
// don't actually do this to get the high-score. it would be sad.
function get_the_cor() {
    // Get points from html
    var points = document.getElementsByClassName("nv-group")[0]

    // create dataset from transform attribute on the svg points
    var transs = Array.from(points.children).map(p => p.getAttribute("transform"))
    var data   = transs.map(x => x.split(",").map(y => y.match(/[\d\.]+/g)).map(Number))
    var tdata  = data[0].map((col, i) => data.map(row => row[i]));

vankesteren

# Pretty print list structures with the unicode stuff
# using https://stackoverflow.com/questions/1649027/how-do-i-print-out-a-tree-structure
str2 <- function(x, name = NULL, max.depth = 10L, indent = NULL, depth = 0L, last = TRUE) {
  cat(indent)
  if (last) {
    cat("\u2514\u2500")
    indent <- paste0(indent, "   ")
  } else {
    cat("\u251c\u2500")
    indent <- paste0(indent, "\u2502  ")

vankesteren

# Gradient descent with autodiff for linear regression
using Zygote

# Data
X = randn(1000, 10)
b = (1:10)
y = X * b + randn(1000)

# MSE for linear model
function mse(bhat)

vankesteren

# convolutions
convolve <- function(x, kernel) {
  size  <- length(kernel)
  x_pad <- c(rep(0, size - 1), x)
  out   <- rep(0, length(x))
  for (i in 1:length(x)) {
    out[i] <- x_pad[i:(i + (size - 1))] %*% kernel
  }
  return(out)
}

vankesteren

# Randomly generate some data
S <- rWishart(1, 10, diag(10))[,,1] / 10
X <- MASS::mvrnorm(100, rep(0, 10), S)
b <- runif(10, -1, 1)
y <- X %*% b + rnorm(100, sd = sqrt(b %*% S %*% b))


# all the interesting regression quantities!
n    <- nrow(X)                       # sample size
p    <- ncol(X)                       # parameters in model matrix

vankesteren

# Use the great rayshader package (www.rayshader.com)
library(rayshader)

# Download data
zip_loc <- tempfile() 
download.file("http://geodata.nationaalgeoregister.nl/ahn1/extract/ahn1_100m/ahn1_100.tif.zip", 
              zip_loc)
local_tif <- raster::raster(unzip(zip_loc, "ahn_100.tif"))
unlink(zip_loc)