Patrick Butler, machine learning researcher

Advent of Code 2024

2024-12-02T00:00:00-05:00

Time permitting I am trying to keep up with the Advent of Code 2024. I will be posting my solutions here as blog entries. I used to really enjoy participating as team captain of one of Virginia Tech's teams for the ACM ICPC. Of course these days I have a full time job and a family so I'll post solutions as I have time or do a few days at a time. I'll be using Python for my solutions as that's the lingua franca of the ML community. My philosophy is to write clean, code that is easy to read and understand. As an AI/ML researcher, my preference will be to use Pandas, then Numpy, then old fashion python as much as it makes sense. Since I am already a day behind I'll just go ahead and post my solutions for the first two days. They're nothing difficult and were enjoyable to write.

Day 1

Day 1 Problem Set

Problem 1

The first one was a simple problem of sorting the data in two list and then finding the absolute different between the lists.

import pandas as pd

data = pd.read_csv("input/1_1.txt", delimiter="\\s+", header=None, index_col=None)

data[0] = data[0].sort_values(ignore_index=True)
data[1] = data[1].sort_values(ignore_index=True)
print((data[0] - data[1]).abs().sum())

Problem 2

The second one was also quite simple. It was a similarity score (though not a distance score! as $\mathrm{score}(a,b) \ne \mathrm{score}(b, a)$ I don't believe is true. As it was, it was a good use of counting the values in one list and then multiplying them by the number of matching values in the other list. Great place to use a dict to count once and lookup many times.

import pandas as pd

data = pd.read_csv("input/1_1.txt", delimiter="\\s+", header=None, index_col=None)

left = data[0]
right = dict(data[1].value_counts())
print(left.apply(lambda x: x * right.get(x, 0)).sum())

Day 2

Day 2 Problem Set

Problem 1

They're already forcing me out of platitudes written in pandas. Part of me is annoyed but part of me is happy to be forced to write something else. Essentially here were we're looking find whether a list is monotonic and that the differences between the elements are in the range of $[1,3]$

import numpy as np

safe = 0
with open("input/2_1.txt") as fp:
    for line in fp:
        report = np.array([int(x) for x in line.strip().split()])
        diff = np.diff(report)
        # check if the signs are the same (add up to length) and 
        # the differences are in the range are <= 3
        if (abs(np.sign(diff).sum()) == len(diff)) & np.all(np.abs(diff) <= 3):
            safe += 1
print(safe)

Problem 2

The second problem took a little more thought. The easiest thing to do was to check if the reactor was safe and if not then drop one of the elements, check again, and iterate until get true or run out of elements to drop. There's no fun in that solution even though these are small problems. The trick to this one is to know that you only have to check where it fails to be safe and drop one of the surrounding values. There is a special case where the first difference is the wrong one and we need to drop the first value in the list, other than that it works mostly the same. I use numpy here because vectorization makes things go zoom and to save on code.

import numpy as np


def is_safe(report, dropped=False):
    diff = np.diff(report)
    zeros = np.argwhere(diff == 0).flatten()
    signs = np.sign(diff)
    monotonic = np.argwhere(signs != signs[0]).flatten()
    # edge case: if everything is wrong then the first element is the problem
    if len(monotonic) == len(signs) - 1:
        monotonic = [0]
    limit = np.argwhere(np.abs(diff) > 3).flatten()
    problems = np.unique(np.concatenate((zeros, monotonic, limit)))
    if len(problems) == 0:
        return True
    if not dropped:
        for i in problems:
            for j in (i, i + 1):
                if is_safe(np.delete(report, j), True):
                    return True
    return False


safe = 0
with open("input/2_1.txt") as fp:
    for line in fp:
        report = np.array([int(x) for x in line.strip().split()])
        if is_safe(report):
            safe += 1
print(safe)

As I am writing this blog post, I realized there is a "mathier" way of doing this: The he distance (or difference if you will) of three points in a list must be interrelated. That is, if $a, b, c$ are in a list, then:

$$ (a - b) + (b - c) = a - c$$

This would allow you to not have to re-calculate the differences each time.

Testing Pelican

2024-11-03T10:20:00-05:00

This is a simple test of the new static generator.

for i in range(10):
  print(i)

Website Refresh

2024-11-01T00:00:00-04:00

Once again I am redoing my website, this time in Pelican because the old static generator is no longer maintained... Realistically I could go ahead and run it un-maintained but there is always the drive for the latest greatest even if you're trying to go the easy route.

Cool Python Trick -- Access repeating numbers

2014-02-11T20:44:00-05:00

If you want to access a list of repeating numbers and the period is different from the array size you can avoid having to calculate for the edge case by using itertools like so:

import numpy as np
import itertools

data = np.arange(10)
endless_data = itertools.cycle(data)
# grab 7 from the array repeatedly,
np.fromiter(islice(a, 7), np.int)
# >>> array([0, 1, 2, 3, 4, 5, 6])

np.fromiter(islice(a, 7), np.int)
# >>> array([7, 8, 9, 0, 1, 2, 3])

Effortless Beamer Slideshows

2013-06-04T12:13:00-04:00

I have always loved having my slides, for various talks, presentations, and lectures in an open and easy to use format and frankly hated using powerpoint. Powerpoint tends to bog down, have a horrible interface and generally suck out my soul. The downside is beamer, the standard slideshow package of LaTeX tends to be bulky requiring many many more lines of code than content, something LaTeX tries to avoid. The answer is PanDoc. PanDoc is a system for generating documents in one format and outputting them in another. It can also do slidewhows outputed as a beamer tex file or as one of several HTML slide show formats.

Workflow

Mostly I use a very similar workflow to what I use for LaTeX. I use a makefile to compile, a PDF viewer with autoupdate to view the output (Skim.app on OSX), and vim to write the document.

My makefile looks like this:

all:
        pandoc -t beamer --template=default.beamer --variable fontsize=10pt -V theme=Darmstadt  presentation.md  -o output.pdf

Where the -t options specifies the output format, and the rest is mostly obvious. The one big change I made was using the following template file: default.beamer. Which removes the autoresizing of images from the template so that I can use inline LaTe X to fine tune image sizes.

Then a slide from my presentation looks like:

    ## Time series analysis
    - Timeseries are much simpler than a full image and therefore easier to work on computationally
    - We generate a time series based upon relevant features
        - Shape
        - Luma
        - Color
    \begin{center}
    \includegraphics[width=.8\textwidth]{../figures/fig-ts1.pdf}
    \end{center}

You can still use the markdown code for images but I find it's usually best to finetune it. Note that this will break compatibility for the html output but I don't use that.

Renewing the Website

2011-10-25T17:53:00-04:00

Why?

My server recently got hacked and while I was pretty sure I had disinfected the system, the only way to truly be sure is to reformat and bring files over after checking them. This meant that it would take me some amount of work to get all the pieces back up and working. Instead, I went static.

Well in the wonderful world of 2.0 there's a lot that can be done using static HTML and some javascript/AJAX. Comments can be outsourced and be done better by people who have the time and inclination to get it done. At some point there will be disqus comments if I make blog posts that people want to comment on.

Static is simpler, I don't have to worry if a fcgi process has failed and I only have one config file to play with (in theory). It also takes less resources.

I had an idea many years ago to create a website using a Makefile and python scripts to glue headers and footers together on webpages. I never implemented it, but now there seems to be many scripts to solve that problem. The particular one I am using is Hyde which is written in Python and has nice templating system as well as several useful builtin plugins.

In any case all I need is a simple website with very small requirements to upkeep. So static, generated sites seem like the best solution.