Last week: Split-apply-combine and plyr

Part I

Debugging basics

Bug!

The original name for glitches and unexpected defects: dates back to at least Edison in 1876, but better story from Grace Hopper in 1947:

(From Wikipedia)

Debugging: what and why?

Debugging is a the process of locating, understanding, and removing bugs from your code

Why should we care to learn about this?

Debugging: how?

Debugging is (largely) a process of differential diagnosis. Stages of debugging:

  1. Reproduce the error: can you make the bug reappear?
  2. Characterize the error: what can you see that is going wrong?
  3. Localize the error: where in the code does the mistake originate?
  4. Modify the code: did you eliminate the error? Did you add new ones?

Reproduce the bug

Step 0: make if happen again

Characterize the bug

Step 1: figure out if it’s a pervasive/big problem

Localize the bug

Step 2: find out exactly where things are going wrong

Localizing can be easy or hard

Sometimes error messages are easier to decode, sometimes they’re harder; this can make locating the bug easier or harder

my.plotter = function(x, y, my.list=NULL) {
  if (!is.null(my.list)) 
    plot(my.list, main="A plot from my.list!")
  else
    plot(x, y, main="A plot from x, y!")
}
my.plotter(my.list=list(x=-10:10, y=(-10:10)^3))

my.plotter() # Easy to understand error message
## Error in plot(x, y, main = "A plot from x, y!"): argument "x" is missing, with no default
my.plotter(my.list=list(x=-10:10, Y=(-10:10)^3)) # Not as clear
## Error in xy.coords(x, y, xlabel, ylabel, log): 'x' is a list, but does not have components 'x' and 'y'

Who called xy.coords()? (Not us, at least not explicitly!) And why is it saying ‘x’ is a list? (We never set it to be so!)

traceback()

Calling traceback(), after an error: traces back through all the function calls leading to the error

If you run my.plotter(my.list=list(x=-10:10, Y=(-10:10)^3)) in the console, then call traceback(), you’ll see:

> traceback()
5: stop("'x' is a list, but does not have components 'x' and 'y'")
4: xy.coords(x, y, xlabel, ylabel, log)
3: plot.default(my.list, main = "A plot from my.list!")
2: plot(my.list, main = "A plot from my.list!") at #4
1: my.plotter(my.list = list(x = -10:10, Y = (-10:10)^3))

We can see that my.plotter() is calling plot() is calling plot.default() is calling xy.coords(), and this last function is throwing the error

Why? Its first argument x is being set to my.list, which is OK, but then it’s expecting this list to have components named x and y (ours are named x and Y)

Part II

Debugging tools

cat(), print()

Most primitive strategy: manually call cat() or print() at various points, to print out the state of variables, to help you localize the error

This is the “stone knives and bear skins” approach to debugging; it is still very popular among some people (actual quote from stackoverflow):

I’ve been a software developer for over twenty years … I’ve never had a problem I could not debug using some careful thought, and well-placed debugging print statements. Many people say that my techniques are primitive, and using a real debugger in an IDE is much better. Yet from my observation, IDE users don’t appear to debug faster or more successfully than I can, using my stone knives and bear skins.

Specialized tools for debugging

R provides you with many debugging tools. Why should we use them, and move past our handy cat() or print() statements?

Let’s see what our primitive hunter found on stackoverflow, after a receiving bunch of comments in response to his quote:

Sweet! … Very illuminating. Debuggers can help me do ad hoc inspection or alteration of variables, code, or any other aspect of the runtime environment, whereas manual debugging requires me to stop, edit, and re-execute.

browser()

One of the simplest but most powerful built-in debugging tools: browser(). Place a call to browser() at any point in your function that you want to debug. As in:

my.fun = function(arg1, arg2, arg3) {
  # Some initial code 
  browser()
  # Some final code
}

Then redefine the function in the console, and run it. Once execution gets to the line with browser(), you’ll enter an interactive debug mode

Things to do while browsing

While in the interactive debug mode granted to you by browser(), you can type any normal R code into the console, to be executed within in the function environment, so you can, e.g., investigate the values of variables defined in the function

You can also type:

(To print any variables named n, s, f, c, or Q, defined in the function environment, use print(n), print(s), etc.)

Browsing in R Studio

You have buttons to click that do the same thing as “n”, “s”, “f”, “c”, “Q” in the “Console” panel; you can see the locally defined variables in the “Environment” panel; the traceback in the “Traceback” panel

Knitting and debugging

As with cat(), print(), traceback(), used for debugging, you should only run browser() in the console, never in an Rmd code chunk that is supposed to be evaluated when knitting

But, to keep track of your debugging code (that you’ll run in the console), you can still use code chunks in Rmd, you just have to specify eval=FALSE

# As an example, here's a code chunk that we can keep around in this Rmd doc,
# but that will never be evaluated (because eval=FALSE) in the Rmd file, take 
# a look at it!
big.mat = matrix(rnorm(1000)^3, 1000, 1000)
big.mat
# Note that the output of big.mat is not printed to the console, and also
# that big.mat was never actually created! (This code was not evaluated)

Part III

Testing

What is testing?

Testing is the systematic writing of additional code to ensure your functions behaves properly.

There’s a lot of topics related to testing, but we’ll focus on two specific aspects: assertions (i.e, ensuring conditions of your functions are satisfied) and unit tests (i.e., simple tests to ensure your function works).

Of course, this requires you to spend more time upfront to write all these things, but it can dramatically save the amount of time you do bug-fixing afterwards.

Assertions

Assertions are boolean checks to ensure that the inputs to your function are properly formatted. For example, if your function expects a matrix, your first few lines of your function should check it is actually a matrix (as opposed to a vector, for example).

If these boolean checks do not pass (i.e, it fails), then you can have the assertion print out a meaningful custom message to pass to the user.

Ideally, if all your assertions pass, your function should never crash afterwards since “everything is as expected”.

assert_that()

We use the assert_that() function in the assertthat package to make assertions. The main benefit to using assert_that() is that you can write custom, meaningful error messages.

Example 1: Function that creates an n by n matrix with 0’s.

#not meaningful errors
create.matrix.simple = function(n){
  matrix(0, n, n)
}

create.matrix.simple(4)
##      [,1] [,2] [,3] [,4]
## [1,]    0    0    0    0
## [2,]    0    0    0    0
## [3,]    0    0    0    0
## [4,]    0    0    0    0
create.matrix.simple(4.1)
##      [,1] [,2] [,3] [,4]
## [1,]    0    0    0    0
## [2,]    0    0    0    0
## [3,]    0    0    0    0
## [4,]    0    0    0    0
create.matrix.simple("asdf")
## Error in matrix(0, n, n): non-numeric matrix extent
library(assertthat)

#meaningful errors
create.matrix = function(n){
  assert_that(length(n) == 1 && is.numeric(n) && n > 0 && n %% 1 == 0, msg = "n is not a positive integer")
  matrix(0, n, n)
}

create.matrix(4)
##      [,1] [,2] [,3] [,4]
## [1,]    0    0    0    0
## [2,]    0    0    0    0
## [3,]    0    0    0    0
## [4,]    0    0    0    0
create.matrix(4.1)
## Error: n is not a positive integer
create.matrix("asdf")
## Error: n is not a positive integer

Example 2: Function that does a linear regression.

mat = matrix(rnorm(20), 10, 2)
colnames(mat) = paste0("X", 1:2)
dat = as.data.frame(mat)

#not meaningful errors
run.lm.simple = function(dat){
  res = lm(X1 ~ ., data = dat)
  coef(res)
}

run.lm.simple(dat)
##  (Intercept)           X2 
## -0.003161468 -0.409245701
run.lm.simple(mat)
## Error in model.frame.default(formula = X1 ~ ., data = dat, drop.unused.levels = TRUE): 'data' must be a data.frame, not a matrix or an array
#meaningful errors
run.lm = function(dat){
  assert_that(is.data.frame(dat), msg = "dat must be a data frame")
  
  res = lm(X1 ~ ., data = dat)
  coef(res)
}

run.lm(dat)
##  (Intercept)           X2 
## -0.003161468 -0.409245701
run.lm(mat)
## Error: dat must be a data frame

Unit testing

Assertions are “tests done on the fly” to ensure that your function can perform properly with the given inputs.

Unit tests, on the other hand, are a suite of tests that your code needs to (or, at least, should) pass at every step of development.

It would seem obvious that when fixing bugs you would want to set up a system of checks that would help to ensure that the bugs do not come back, or that other bugs are not introduced when updating code. But what is less obvious is how to do that.

Often times, you can never fully ensure that your code will always behave properly on any input. However, if your function behaves properly on a wide range of simple inputs where you know what the intended outcome should be, you can extrapolate comfortably to know that, for the most part, your function will behave properly for most inputs.

test_that()

We use the test_that() function in the testthat package to make unit tests. This allows a clean format to write tests that you can document.

Each test consists of two parts: 1) a message that describes what you are testing, and 2) code to execute that results in a TRUE or FALSE. If TRUE, your test passed. If FALSE (or your testing code crashed), your test failed and you should investigate your function and testing code to see why your test failed.

Typically, you’ll write tests for simple inputs for your function such that you know exactly what the output should be.

Most commonly, you will use expect_true() or expect_error() as the last line of your test.

Example 1: Given a matrix mat and row indices idx, compute the median of each column of mat[idx,].

library(testthat)
## 
## Attaching package: 'testthat'
## The following object is masked from 'package:devtools':
## 
##     setup
#the incorrect version
col.median = function(mat, idx){
  apply(mat[idx,], 2, median)
}

mat = matrix(1:16, 4, 4)
mat
##      [,1] [,2] [,3] [,4]
## [1,]    1    5    9   13
## [2,]    2    6   10   14
## [3,]    3    7   11   15
## [4,]    4    8   12   16
test_that("col.median() works", {
  res = col.median(mat, 1:3)
  
  expect_true(all(res == c(2,6,10,14)))
})

test_that("col.median() works for one row", {
  res = col.median(mat, 1)
  
  expect_true(all(res == c(1,5,9,13)))
})
## Error: Test failed: 'col.median() works for one row'
## * dim(X) must have a positive length
## 1: col.median(mat, 1) at <text>:18
## 2: apply(mat[idx, ], 2, median) at <text>:5
## 3: stop("dim(X) must have a positive length")
test_that("col.median() errors for non-integer indices", {
  expect_error(col.median(mat, c(1.4, 2)))
})
## Error: Test failed: 'col.median() errors for non-integer indices'
## * `col.median(mat, c(1.4, 2))` did not throw an error.
#the correct version
col.median = function(mat, idx){
  assert_that(all(idx %% 1 == 0), msg = "idx must be all integers")
  
  apply(mat[idx,,drop=FALSE], 2, median)
}

test_that("col.median() works", {
  res = col.median(mat, 1:3)
  
  expect_true(all(res == c(2,6,10,14)))
})

test_that("col.median() works for one row", {
  res = col.median(mat, 1)
  
  expect_true(all(res == c(1,5,9,13)))
})

test_that("col.median() errors for non-integer indices", {
  expect_error(col.median(mat, c(1.4, 2)))
})

Example 2: Given a numeric vector, find the first index (from left to right) of the number 2. Then add 1 to all values from that index to the end (right) of the vector.

#incorrect version
increment = function(vec){
  n = length(vec)
  idx = which(vec == 2)
  vec[idx[1]:n] = vec[idx[1]:n]+1
  vec
}

test_that("increment() works", {
  res = increment(1:10)
  
  expect_true(all(res == c(1,3:11)))
})

test_that("increment() works when no 2 is in the vector", {
  res = increment(3:10)
  
  expect_true(all(res == 3:10))
})
## Error: Test failed: 'increment() works when no 2 is in the vector'
## * NA/NaN argument
## 1: increment(3:10) at <text>:16
#correct version
increment = function(vec){
  n = length(vec)
  idx = which(vec == 2)
  if(length(idx) == 0) return(vec)
  
  vec[idx[1]:n] = vec[idx[1]:n]+1
  vec
}

test_that("increment() works", {
  res = increment(1:10)
  
  expect_true(all(res == c(1,3:11)))
})

test_that("increment() works when no 2 is in the vector", {
  res = increment(3:10)
  
  expect_true(all(res == 3:10))
})

Unit testing: how?

Unit testing is essentially a log of all the properties of your function that you want remember to check. Since you’ll be changing your code constantly, you want to be assured that when “fixing one bug”, “an old bug that used to be fixed does not reappear”.

Useful tips of unit testing:

Summary