Compsci 100, Spring 2010 18.1 What’s left to talk about? l Transforms Making Huffman compress...

Compsci 100, Spring 2010 18.1

What’s left to talk about?

Transforms Making Huffman compress more Understanding what transforms do

• Conceptual understanding, details left to …• All information here, we won’t discuss details

Expressing concepts in different languages How hard is it to learn C++, Python, … Are there “different” languages?

• Ruby, Scheme, …


What is a transform?

Multiply two near-zero numbers, what happens? Add their logarithms: log(a)+log(b) = log(ab), invertible What is log of 10-13? Benefits of transform?

What is FFT: Fast Fourier Transform? O(n log n) method for computing a Fourier Transform Better than O(n2), huge difference for lots of data

points Shazam? how shazam might work

Feature extraction from images: faces, edges, lines, … Hough transform

Wavelet transforms do something too, but … http://en.wikipedia.org/wiki/Ingrid_Daubechies


Burrows Wheeler Transform

Michael Burrows and David Wheeler in 1994, BWT By itself it is NOT a compression scheme

It’s used to preprocess data, or transform data, to make it more amenable to compression like Huffman Coding

Huff depends on redundancy/repetition, as do many compression schemes

http://en.wikipedia.org/wiki/Burrows-Wheeler_transform http://marknelson.us/1996/09/01/bwt

Main idea in BWT: transform the data into something more compressible and make the transform fast, though it will be slower than no transform TANSTAAFL (what does this mean?)


David Wheeler (1927-2004)

Invented subroutine “Wheeler was an

inspiring teacher who helped to develop computer science teaching at Cambridge from its inception in 1953, when the Diploma in Computer Science was launched as the world's first taught course in computing.”


Mike Burrows

He's one of the pioneers of the information age. His invention of Alta Vista helped open up an entire new route for the information highway that is still far from fully explored. His work history, intertwined with the development of the high-tech industry over the past two decades, is distinctly a tale of scientific genius.

http://www.stanford.edu/group/gpj/cgi-bin/drupal/?q=node/60


BWT efficiency

BWT is a block transform – requires storing n copies of the file with time O(n log n) to sort copy (file has length n) We can’t really do this in practice in terms of storage Instead of storing n copies of the file, store one copy

and an integer index (break file into blocks of size n) But sorting is still O(n log n) and it’s actually worse

Each comparison in the sort looks at the entire file In normal sort analysis the comparison is O(1),

strings are small Now we have key comparison of O(n), so sort is

actually…

O(n2 log n), why?


BWT at 10,000 ft: big picture

Remember, goal is to exploit/create repetition (redundancy) Create repetition as follows Consider original text: duke blue devils. Create n copies by shifting/rotating by one

character0: duke blue devils.1: uke blue devils.d2: ke blue devils.du3: e blue devils.duk4: blue devils.duke5: blue devils.duke 6: lue devils.duke b7: ue devils.duke bl8: e devils.duke blu

9: devils.duke blue10: devils.duke blue 11: evils.duke blue d12: vils.duke blue de13: ils.duke blue dev14: ls.duke blue devi15: s.duke blue devil16: .duke blue devils


BWT at 10,000 ft: big picture

Once we have n copies (but not really n copies!) Sort the copies Remember the comparison will be O(n) We’ll look at the last column, see next slide

• What’s true about first column?

4: blue devils.duke9: devils.duke blue16: .duke blue devils5: blue devils.duke 10: devils.duke blue 0: duke blue devils.3: e blue devils.duk8: e devils.duke blu11: evils.duke blue d

13: ils.duke blue dev2: ke blue devils.du14: ls.duke blue devi6: lue devils.duke b15: s.duke blue devil7: ue devils.duke bl1: uke blue devils.d12: vils.duke blue de


|ees .kudvuibllde| | .bddeeeikllsuuv|4: blue devils.duke9: devils.duke blue16: .duke blue devils5: blue devils.duke 10: devils.duke blue 0: duke blue devils.3: e blue devils.duk8: e devils.duke blu11: evils.duke blue d13: ils.duke blue dev2: ke blue devils.du14: ls.duke blue devi6: lue devils.duke b15: s.duke blue devil7: ue devils.duke bl1: uke blue devils.d12: vils.duke blue de

Properties of first column Lexicographical order Maximally ‘clumped’ why? From it, can we create last?

Properties of last column Some clumps (real files) Can we create first? Why?

See row labeled 8: Last char precedes first in

original! True for all rows! Can recreate everything:

Simple (code) but hard (idea)


What do we know about last column?

Contains every character of original file Why is there repetition in the last column? Is there repetition in the first column?

Keep the last column because we can recreate the first What’s in every column of the sorted list? If we have the last column we can create the first

• Sorting the last column yields first We can create every column which means if we

know what row the original text is in we’re done!• Look back at sorted rows, what row has index

0?


BWT from a 5,000 ft view

How do we avoid storing n copies of the input file? Store once with index of what the first character

is 0 and “duke blue devils.” is the original string 3 and “duke blue devils.” is “e blue devils. du” What is 7 and “duke blue devils.”

You’ll be given a class Rotatable that can be sorted Construct object from original text and index When compared, use the index as a place to start Rotatable can report the last char of any “row” Rotatable can report its index (stored on

construction)


BWT 2,000 feet

To transform all we need is the last column and the row at which the original string is in the list of sorted strings We take these two pieces of information and

either compress them or transform them further

After the transform we run Huff on the result

We can’t store/sort a huge file, what do we do? Process big files in chunks/blocks

• Read block, transform block, Huff block• Read block, transform block, Huff block…• Block size may impact performance


Toward BWT from zero feet

First look at code for HuffProcessor.compress Tree already made, preprocessCompress How writeHeader,writeCompressedData work?

public int compress(InputStream in, OutputStream out) {BitOutputStream bout = new BitOutputStream(out);BitInputStream bin = new BitInputStream(in); int bitCount = 0;

myRoot = makeTree(); makeMapEncodings(myRoot,””); bitCount += writeHeader(bout); bitCount += writeCompressedData(bin,bout); bout.flush(); return bitCount;}


BWT from zero feet, part I

Read a block of data, transform it, then huff it To huff we write a magic number, write header/tree,

and write compressed bits based on Huffman encodings

We already have huff code, need to use on a transformed bunch of characters rather than on the input file

So process input stream by passing it to BW transform which reads a chunk and returns char[], the last column A char is a 16-bit, unsigned value, we only need 8-

bit value, but use char because we can’t use byte• In Java byte is signed, -128,.. 127• What does all that mean?


Use what we have, need new stream

We want to use existing compression code we wrote before Read a block of 8-bit/chunks, store in char[] array Repeat until no more blocks, last block not full? Block as char[], treat as stream and feed it to Huff

• Count characters, make tree, compress

We need an Adapter, something that takes char[] array and turns it into an InputStream which we feed to Huff compressor ByteArrayInputStream, turns byte[] to stream We can store 8-bit chunks as bytes for stream

purposes


ByteArrayInputStream and blockspublic int compress(InputStream in, OutputStream out) {

BitOutputStream boout = new BitOutputStream(out); BitInputStream bin = new BitInputStream(in);

int bitCount = 0;

BurrowsWheeler bwt = new BurrowsWheeler(); while (true){ char[] chunk = bw.transform(bin); if (chunk.length < 1) break; chunk = btw.mtf(chunk); byte[] array = new byte[chunk.length]; for(int k=0; k < array.length; k++){ array[k] = (byte) chunk[k]; } ByteArrayInputStream bas = new ByteArrayInputStream(array); preprocessInitialize(bas); myRoot = makeTree(); makeMapEncodings(myRoot,””); BitInputStream blockBis = new BitInputStream(new ByteArrayInputStream(array)); bitCount += writeHeader(bout); bitCount += writeCompressedData(blockBis,bout);

} bout.flush(); return bitCount;}


How do we untransform?

Untransforming is very slick Basically sort the last column in O(n) time Run an O(n) algorithm to get back original

block

We sort the last column in O(n) time using a counting sort, which is sometimes one phase of radix sort Call sort: easier to code and a good first step The counting sort leverages that we’re sorting

“characters” --- whatever we read when doing compression which is an 8-bit chunk

How many different 8-bit chunks are there?


Counting sort

If we have an array of integers all of whose values are between 0 and 255, how can we sort by counting number of occurrences of each integer? Suppose we have 4 occurrences of one, 1

occurrence of two, 3 occurrences of five and 2 occurrences of seven, what’s the sorted array? (we don’t know the original, just the counts)

What’s the answer? How do we write code to do this?

More than one way, as long as O(n) doesn’t matter really


Another transform: Move To Front In practice we can introduce more repetition and

redundancy using a Move-to-front transform (MTF) We’re going to compress a sequence of numbers (the

8-bit chunks we read, might be the last column from BWT)

Instead of just writing the numbers, use MTF to write

Introduce more redundancy/repetition if there are runs of characters. For example: consider “AAADDDFFFF” As numbers this is 97 97 97 100 100 100 102 102 102 Using MTF, start with index[k] = k

• 0,1,2,3,4,…,96,97,98,99,…,255 Search for 97, initially it’s at index[97], then MTF

• 97,0,1,2,3,4,5,…, 96,98,99,…,255


More on why MTF works As numbers this is 97 97 97 100 100 100 102 102 102

Using MTF, start with index[k] = k Search for 97, initially it’s at index[97], then MTF

• 97,0,1,2,3,4,5,…,96,98,99,100,101,… Next time we search for 97 where is it? At 0!

So, to write out 97 97 97 we actually write 97 0 0, then we write out 100, where is it? Still at 100, why? Then MTF: 100,97,0,1,2,3,…96,98,99,101,102,…

So, to write out 97 97 97 100 100 100 102 102 102 we write: 97, 0, 0, 100, 0, 0, 102, 0, 0 Lots of zeros, ones, etc. Thus more Huffable, why?


Complexity of MTF and UMTF

Given n characters, we have to look through 256 indexes (worst case) So, 256*n, this is …. O(n) Average case is much better, the whole point of

MTF is to find repeats near the beginning (what about MTF complexity?)

How to untransform, undo MTF, e.g., given 97, 0, 0, 100, 0, 0, 102, 0, 0

How do we recover AAADDDFFF (97,97,97,100,100,…102) Initially index[k] = k, so where is 97? O(1) look

up, then MTF


Burrows Wheeler Summary

Transform data: make it more “compressable” Introduce redundancy First do BWT, then do MTF (latter provided) Do this in chunks For each chunk array (after BWT and MTF) huff

it

To uncompress data Read block of huffed data, uncompress it,

untransform Undo MTF, undo BWT: this code is given to you Don’t forget magic numbers


John Tukey: 1915-2000

Cooley-Tukey FFT Bit: Binary Digit Box-plot “software” used in print

Far better an approximate answer to the right question, which is often vague, than an exact answer to the wrong question, which can always be made precise.

The combination of some data and an aching desire for an answer does not ensure that a reasonable answer can be extracted from a given body of data.


A Rose by any other name…C or Java?

Why do we use Java in our courses (royal we?) Object oriented Large collection of libraries Safe for advanced programming and

beginners Harder to shoot ourselves in the foot

Why don't we use C++ (or C)? Standard libraries weak or non-existant

(comparatively) Easy to make mistakes when beginning No GUIs, complicated compilation model What about other languages?


Why do we learn other languages? Perl, Python, PHP, Ruby, C, C++, Java, Scheme,

ML, Can we do something different in one

language?• Depends on what different means.• In theory: no; in practice: yes

What languages do you know? All of them. In what languages are you fluent? None of them

In later courses why do we use C or C++? Closer to the machine, understand abstractions

at many levels Some problems are better suited to one

language•Writing an operating system? Linux?


Unique words in Java

import java.util.*;

import java.io.*;public class Unique { public static void main(String[] args) throws IOException{ Scanner scan = new Scanner(new File("/data/melville.txt")); TreeSet<String> set = new TreeSet<String>(); while (scan.hasNext()){ String str = scan.next(); set.add(str); } for(String s : set){ System.out.println(s); } }}


Bjarne Stroustrup, Designer of C++

Numerous awards, engineering and science ACM Grace Hopper

Formerly at Bell Labs Now Texas A&M

“There's an old story about the person who wished his computer was as easy to use as his telephone. That wish has come true, since I no longer know how to use my telephone.”

Bjarne Stroustrup


Unique words in C++#include <iostream>#include <fstream>#include <set>using namespace std;

int main(){ ifstream input("/data/melville.txt"); set<string> unique; string word; while (input >> word){ unique.insert(word); } set<string>::iterator it = unique.begin(); for(; it != unique.end(); it++){ cout << *it << endl; } return 0;}


PHP, Rasmus Lerdorf and Others

Rasmus Lerdorf Qeqertarsuaq, Greenland 1995 started PHP, now part

of it http://en.wikipedia.org/

wiki/PHP Personal Home Page

No longer an acronym

“When the world becomes standard, I will start caring about standards.”

Rasmus Lerdorf


Unique words in PHP<?php

$wholething = file_get_contents("file:///data/melville.txt");$wholething = trim($wholething);

$array = preg_split("/\s+/",$wholething);$uni = array_unique($array);sort($uni);foreach ($uni as $word){ echo $word."<br>";}

?>


Guido van Rossum

BDFL for Python development Benevolent Dictator For Life Late 80’s began development

Python is multi-paradigm OO, Functional, Structured, …

We're looking forward to a future where every computer user will be able to "open the hood" of their computer and make improvements to the applications inside. We believe that this will eventually change the nature of software and software development tools fundamentally.

Guido van Rossum, 1999!


Unique Words in Python

#! /usr/bin/env python

import sysimport re

def main(): f = open('/data/melville.txt', 'r') words = re.split('\s+',f.read().strip()) allWords = set() for w in words: allWords.add(w) for word in sorted(allWords): print "%s" % word

if __name__ == "__main__": main()


Kernighan and Ritchie

First C book, 1978 First ‘hello world’ Ritchie: Unix too!

Turing award 1983 Kernighan: tools

Strunk and White

Everyone knows that debugging is twice as hard as writing a program in the first place. So if you are as clever as you can be when you write it, how will you ever debug it?

Brian Kernighan


How do we read a file in C?

#include <stdio.h>#include <string.h>#include <stdlib.h>

int strcompare(const void * a, const void * b){ char ** stra = (char **) a; char ** strb = (char **) b; return strcmp(*stra, *strb);}

int main(){ FILE * file = fopen("/data/melville.txt","r"); char buf[1024]; char ** words = (char **) malloc(5000*sizeof(char **)); int count = 0; int k;


Storing words read when reading in C

while (fscanf(file,"%s",buf) != EOF){ int found = 0; // look for word just read

for(k=0; k < count; k++){ if (strcmp(buf,words[k]) == 0){ found = 1; break; } } if (!found){ // not found, add to list

words[count] = (char *) malloc(strlen(buf)+1); strcpy(words[count],buf); count++; } }

Complexity of reading/storing? Allocation of memory?


Sorting, Printing, Freeing in C

qsort(words,count,sizeof(char *), strcompare); for(k=0; k < count; k++) { printf("%s\n",words[k]); }

for(k=0; k < count; k++){ free(words[k]); } free(words);

} Sorting, printing, and freeing

How to sort? What’s analgous to comparator?

Why do we call free? Necessary in this program? Why?

Compsci 100, Spring 2010 18.1 What’s left to talk about? l Transforms Making Huffman compress...

Documents

Transcript of Compsci 100, Spring 2010 18.1 What’s left to talk about? l Transforms Making Huffman compress...