Related
I want to change the outerstride of Eigen matrix at compile time without using map function.
I tried to change this using OuterStrideAtCompileTime variable,but it doesn't work.Is there any way to do that ?
One more thing printing mat.Outerstride() every time gives number of rows of input matrix.How to print OuterStride of eigen matrix?
Thanks in advance.
I was defining an eigen matrix with map function like
MatrixXf mat;
float arr[16] = {1,2,3,4,5,6,7,8,9,10,11,12,13,14,15,16};
mat = Map<Matrix<float,Dynamic,Dynamic,Eigen::RowMajor>, 0, OuterStride<Dynamic> > (arr,4,4,OuterStride<Dynamic>(5));
It's working fine, whenever I tried to change the outer stride by using
mat.OuterStrideAtCompileTime = 7;
It's not working.
Outer stride is a parameter related to the data storage. A more commonly used name is leading dimension. You could find some explaination here.
http://www.ibm.com/support/knowledgecenter/SSFHY8_5.3.0/com.ibm.cluster.essl.v5r3.essl100.doc/am5gr_leaddi.htm
Basically for an existing matrix, it can not be changed. The only way to change it without changing the elements of the matrix is to copy the matrix to a new memory space using a different outer stride setting. This usually happens when you copy a matrix into another as a sub-matrix.
For a column major matrix the minimum possible outer stride equals to the number of rows, as the number you have printed out.
When using Eigen, you don't need worry about it as Eigen usually takes care of it for you expect for Eigen::Map.
You code actually doesn't work. Setting out stride to 5 is already out of range as the existing matrix(4x4) stored in arr is of stride 4 and stride 5 x 4 columns = 20 > 16.
#include <iostream>
#include <Eigen/Eigen>
int main(void) {
using namespace Eigen;
MatrixXf mat;
float arr[16] = { 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16 };
mat = Map<Matrix<float, Dynamic, Dynamic, Eigen::RowMajor>, 0,
OuterStride<Dynamic> >(arr, 4, 4, OuterStride<Dynamic>(5));
std::cout << "mat with stride 5:\n" << mat << std::endl;
mat = Map<Matrix<float, Dynamic, Dynamic, Eigen::RowMajor>, 0,
OuterStride<Dynamic> >(arr, 4, 4, OuterStride<Dynamic>(4));
std::cout << "mat with stride 4:\n" << mat << std::endl;
return 0;
}
Please compare the output.
mat with stride 5:
1 2 3 4
6 7 8 9
11 12 13 14
16 0 0 5.01639e-14
mat with stride 4:
1 2 3 4
5 6 7 8
9 10 11 12
13 14 15 16
If you extend the array to 20 elements
#include <iostream>
#include <Eigen/Eigen>
int main(void) {
using namespace Eigen;
MatrixXf mat;
float arr[20] = { 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20 };
Map<Matrix<float, Dynamic, Dynamic, Eigen::RowMajor>, 0, OuterStride<Dynamic> > map1(arr, 4, 4, OuterStride<Dynamic>(5));
mat = map1;
std::cout << "map1 outer stride: " << map1.outerStride() << std::endl << map1 << std::endl;
std::cout << "mat outer stride: " << mat.outerStride() << std::endl << mat << std::endl;
Map<Matrix<float, Dynamic, Dynamic, Eigen::RowMajor>, 0, OuterStride<Dynamic> > map2(arr, 4, 4, OuterStride<Dynamic>(4));
mat = map2;
std::cout << "map2 outer stride: " << map2.outerStride() << std::endl << map2 << std::endl;
std::cout << "mat outer stride: " << mat.outerStride() << std::endl << mat << std::endl;
return 0;
}
The output will be
map1 outer stride: 5
1 2 3 4
6 7 8 9
11 12 13 14
16 17 18 19
mat outer stride: 4
1 2 3 4
6 7 8 9
11 12 13 14
16 17 18 19
map2 outer stride: 4
1 2 3 4
5 6 7 8
9 10 11 12
13 14 15 16
mat outer stride: 4
1 2 3 4
5 6 7 8
9 10 11 12
13 14 15 16
You could also see the outer stride changes when copying map1 to mat.
Hope this give you a better view of what out stride is.
In fact in your original code, you are using Map in a wrong way - you shouldn't have copied Map() to Matrix mat.
That's why when you print the stride of mat, it is always 4.
What you need to do is to eliminate the unnecessary data copy and print the stride of map1/map2.
I have a color-wheel that maps a color to each hour on a 24-hour clock. Now given the hour of day, I want to map those colors to a 12-hour clock such that the colors 5 hours before and 6 hours after the current hour are used. But it gets a bit tricky b/c the 0th index of the result always has to be the 0th color or the 12th color of the 24 color-wheel.
For example, given colors24 as an array of 24 colors and a hour time of 5 then the final color12 array would map to colors24's indexes as:
{0,1,2,3,4,5,6,7,8,9,10,11}
If the hour is 3, then:
{0,1,2,3,4,5,6,7,8,9,22,23}
And if the hour is 9, then:
{12,13,14,15,4,5,6,7,8,9,10,11}
Bonus points if the algorithm can be generalized to any two arrays regardless of size so long as the first is evenly divisible by the second.
If hours is the total number of hours (24), length the number of colors displayed at a time (12), and hour is the current hour, then this is a generic algorithm to get the indexes into the color array:
result = [];
add = hour + hours - (length / 2) - (length % 2) + 1;
for (i = 0; i < length; i++) {
result[(add + i) % length] = (add + i) % hours;
}
Here is a Javascript implementation (generic, can be used with other ranges than 24/12):
function getColorIndexes(hour, hours, length) {
var i, result, add;
if (hours % length) throw "number of hours must be multiple of length";
result = [];
add = hour + hours - (length / 2) - (length % 2) + 1;
for (i = 0; i < length; i++) {
result[(add + i) % length] = (add + i) % hours;
}
return result;
}
console.log ('hour=3: ' + getColorIndexes(3, 24, 12));
console.log ('hour=5: ' + getColorIndexes(5, 24, 12));
console.log ('hour=9: ' + getColorIndexes(9, 24, 12));
console.log ('hour=23: ' + getColorIndexes(23, 24, 12));
As stated in the question, the number of hours (24) must be a multiple of the length of the array to return.
This can be done by first placing the numbers into a temporary array, then finding the location of 0 or 12 in it, and printing the results from that position on, treating the index as circular (i.e. modulo the array length)
Here is an example implementation:
int num[12];
// Populate the values that we are going to need
for (int i = 0 ; i != 12 ; i++) {
// 19 is 24-5
num[i] = (h+i+19) % 24;
}
int p = 0;
// Find p, the position of 0 or 12
while (num[p] != 0 && num[p] != 12) {
p++;
}
// Print num[] array with offset of p
for (int i = 0 ; i != 12 ; i++) {
printf("%d ", num[(p+i) % 12]);
}
Demo.
Note: The first and the second loops can be combined. Add a check if the number you just set is zero or 12, and set the value of p when you find a match.
Can you not get the colors straight away, i.e. from (C-Y/2+X+1)%X to (C+Y/2)%X, and then sort them?
(This is the same as looping (C+Z+X+1)%X from Z = -Y/2 to Z = Y/2-1):
for (i = 0, j = c+x+1, z = -y/2; z < y/2; z++) {
color[i++] = (z+j)%x;
}
For C=3, X=24 and Y=12, you get:
(C-12/2+24+1)%24 = 3-6+24+1 = 22, 23, 0, 1 .. 9
After sorting you get 0, 1 ...9, 22, 23 as requested.
Without sorting, you'd always get a sequence with the current hour smack in the middle (which could be good for some applications), while your 3 example has it shifted left two places.
You can do this by shifting instead of sorting by noticing that you only need to shift if c is below Y/2 (C=3 makes you start from -2, which becomes 22), in which case you shift by negative y/2-c (here, 2, or 12+2 using another modulus), or if c > (x-y/2), in which case you'd end beyond x: if c = 20, c+6 is 26, which gets rolled back to 2:
15 16 17 18 19 20 21 22 23 0 1 2
and gives a s factor of 2+1 = 3, or (c+y/2)%x+1 in general:
0 1 2 15 16 17 18 19 20 21 22 23
for (i = 0, j = c+x+1, z = -y/2; z < y/2; z++) {
color[(s+i++)%y] = (z+j)%x;
}
However, I think you've got a problem if x > 2*y; in that case you get some c values for which neither 0, nor x/2 are "in reach" of c. That is, "evenly divisible" must then mean that x must always be equal to y*2.
Here is a solution in JavaScript:
function f(h) {
var retval = [];
for (var i = h - 5; i <= h + 6; ++i)
retval.push((i+24) % 24);
return retval.sort(function(a,b){return a-b;}); // This is just a regular sort
}
https://repl.it/CWQf
For example,
f(5) // [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11 ]
f(3) // [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 22, 23 ]
f(9) // [ 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15 ]
I'm interested in finding the nth row of pascal triangle (not a specific element but the whole row itself). What would be the most efficient way to do it?
I thought about the conventional way to construct the triangle by summing up the corresponding elements in the row above which would take:
1 + 2 + .. + n = O(n^2)
Another way could be using the combination formula of a specific element:
c(n, k) = n! / (k!(n-k)!)
for each element in the row which I guess would take more time the the former method depending on the way to calculate the combination. Any ideas?
>>> def pascal(n):
... line = [1]
... for k in range(n):
... line.append(line[k] * (n-k) / (k+1))
... return line
...
>>> pascal(9)
[1, 9, 36, 84, 126, 126, 84, 36, 9, 1]
This uses the following identity:
C(n,k+1) = C(n,k) * (n-k) / (k+1)
So you can start with C(n,0) = 1 and then calculate the rest of the line using this identity, each time multiplying the previous element by (n-k) / (k+1).
A single row can be calculated as follows:
First compute 1. -> N choose 0
Then N/1 -> N choose 1
Then N*(N-1)/1*2 -> N choose 2
Then N*(N-1)*(N-2)/1*2*3 -> N choose 3
.....
Notice that you can compute the next value from the previous value, by just multipyling by a single number and then dividing by another number.
This can be done in a single loop. Sample python.
def comb_row(n):
r = 0
num = n
cur = 1
yield cur
while r <= n:
r += 1
cur = (cur* num)/r
yield cur
num -= 1
The most efficient approach would be:
std::vector<int> pascal_row(int n){
std::vector<int> row(n+1);
row[0] = 1; //First element is always 1
for(int i=1; i<n/2+1; i++){ //Progress up, until reaching the middle value
row[i] = row[i-1] * (n-i+1)/i;
}
for(int i=n/2+1; i<=n; i++){ //Copy the inverse of the first part
row[i] = row[n-i];
}
return row;
}
here is a fast example implemented in go-lang that calculates from the outer edges of a row and works it's way to the middle assigning two values with a single calculation...
package main
import "fmt"
func calcRow(n int) []int {
// row always has n + 1 elements
row := make( []int, n + 1, n + 1 )
// set the edges
row[0], row[n] = 1, 1
// calculate values for the next n-1 columns
for i := 0; i < int(n / 2) ; i++ {
x := row[ i ] * (n - i) / (i + 1)
row[ i + 1 ], row[ n - 1 - i ] = x, x
}
return row
}
func main() {
for n := 0; n < 20; n++ {
fmt.Printf("n = %d, row = %v\n", n, calcRow( n ))
}
}
the output for 20 iterations takes about 1/4 millisecond to run...
n = 0, row = [1]
n = 1, row = [1 1]
n = 2, row = [1 2 1]
n = 3, row = [1 3 3 1]
n = 4, row = [1 4 6 4 1]
n = 5, row = [1 5 10 10 5 1]
n = 6, row = [1 6 15 20 15 6 1]
n = 7, row = [1 7 21 35 35 21 7 1]
n = 8, row = [1 8 28 56 70 56 28 8 1]
n = 9, row = [1 9 36 84 126 126 84 36 9 1]
n = 10, row = [1 10 45 120 210 252 210 120 45 10 1]
n = 11, row = [1 11 55 165 330 462 462 330 165 55 11 1]
n = 12, row = [1 12 66 220 495 792 924 792 495 220 66 12 1]
n = 13, row = [1 13 78 286 715 1287 1716 1716 1287 715 286 78 13 1]
n = 14, row = [1 14 91 364 1001 2002 3003 3432 3003 2002 1001 364 91 14 1]
n = 15, row = [1 15 105 455 1365 3003 5005 6435 6435 5005 3003 1365 455 105 15 1]
n = 16, row = [1 16 120 560 1820 4368 8008 11440 12870 11440 8008 4368 1820 560 120 16 1]
n = 17, row = [1 17 136 680 2380 6188 12376 19448 24310 24310 19448 12376 6188 2380 680 136 17 1]
n = 18, row = [1 18 153 816 3060 8568 18564 31824 43758 48620 43758 31824 18564 8568 3060 816 153 18 1]
n = 19, row = [1 19 171 969 3876 11628 27132 50388 75582 92378 92378 75582 50388 27132 11628 3876 969 171 19 1]
An easy way to calculate it is by noticing that the element of the next row can be calculated as a sum of two consecutive elements in the previous row.
[1, 5, 10, 10, 5, 1]
[1, 6, 15, 20, 15, 6, 1]
For example 6 = 5 + 1, 15 = 5 + 10, 1 = 1 + 0 and 20 = 10 + 10. This gives a simple algorithm to calculate the next row from the previous one.
def pascal(n):
row = [1]
for x in xrange(n):
row = [l + r for l, r in zip(row + [0], [0] + row)]
# print row
return row
print pascal(10)
In Scala Programming: i would have done it as simple as this:
def pascal(c: Int, r: Int): Int = c match {
case 0 => 1
case `c` if c >= r => 1
case _ => pascal(c-1, r-1)+pascal(c, r-1)
}
I would call it inside this:
for (row <- 0 to 10) {
for (col <- 0 to row)
print(pascal(col, row) + " ")
println()
}
resulting to:
.
1
1 1
1 2 1
1 3 3 1
1 4 6 4 1
1 5 10 10 5 1
1 6 15 20 15 6 1
1 7 21 35 35 21 7 1
1 8 28 56 70 56 28 8 1
1 9 36 84 126 126 84 36 9 1
1 10 45 120 210 252 210 120 45 10 1
To explain step by step:
Step 1: We make sure that if our column is the first one we always return figure 1.
Step 2: Since each X-th row there are X number of columns. So we say that; the last column X is greater than or equal to X-th row, then the return figure 1.
Step 3: Otherwise, we get the sum of the repeated pascal of the column just before the current one and the row just before the current one ; and the pascal of that column and the row just before the current one.
Good Luck.
Let me build upon Shane's excellent work for an R solution. (Thank you, Shane!. His code for generating the triangle:
pascalTriangle <- function(h) {
lapply(0:h, function(i) choose(i, 0:i))
}
This will allow one to store the triangle as a list. We can then index whatever row desired. But please add 1 when indexing! For example, I'll grab the bottom row:
pt_with_24_rows <- pascalTriangle(24)
row_24 <- pt_with_24_rows[25] # add one
row_24[[1]] # prints the row
So, finally, make-believe I have a Galton Board problem. I have the arbitrary challenge of finding out percentage of beans have clustered in the center: say, bins 10 to 15 (out of 25).
sum(row_24[[1]][10:15])/sum(row_24[[1]])
Which turns out to be 0.7704771. All good!
In Ruby, the following code will print out the specific row of Pascals Triangle that you want:
def row(n)
pascal = [1]
if n < 1
p pascal
return pascal
else
n.times do |num|
nextNum = ((n - num)/(num.to_f + 1)) * pascal[num]
pascal << nextNum.to_i
end
end
p pascal
end
Where calling row(0) returns [1] and row(5) returns [1, 5, 10, 10, 5, 1]
Here is the another best and simple way to design a Pascal Triangle dynamically using VBA.
`1
11
121
1331
14641`
`Sub pascal()
Dim book As Excel.Workbook
Dim sht As Worksheet
Set book = ThisWorkbook
Set sht = book.Worksheets("sheet1")
a = InputBox("Enter the Number", "Fill")
For i = 1 To a
For k = 1 To i
If i >= 2 And k >= 2 Then
sht.Cells(i, k).Value = sht.Cells(i - 1, k - 1) + sht.Cell(i- 1, k)
Else
sht.Cells(i, k).Value = 1
End If
Next k
Next i
End Sub`
I used Ti-84 Plus CE
The use of –> in line 6 is the store value button
Forloop syntax is
:For(variable, beginning, end [, increment])
:Commands
:End
nCr syntax is
:valueA nCr valueB
List indexes start at 1 so that's why i set it to R+1
N= row
R= column
PROGRAM: PASCAL
:ClrHome
:ClrList L1
:Disp "ROW
:Input N
:For(R,0,N,1)
:N nCr R–>L1(R+1)
:End
:Disp L1
This is the fastest way I can think of to do this in programming (with a ti 84) but if you mean to be able to calculate the row using pen and paper then just draw out the triangle cause doing factorals are a pain!
Here's an O(n) space-complexity solution in Python:
def generate_pascal_nth_row(n):
result=[1]*n
for i in range(n):
previous_res = result.copy()
for j in range(1,i):
result[j] = previous_res[j-1] + previous_res[j]
return result
print(generate_pascal_nth_row(6))
class Solution{
public:
int comb(int n,int r){
long long c=1;
for(int i=1;i<=r;i++) { //calculates n!/(n-r)!
c=((c*n))/i; n--;
}
return c;
}
vector<int> getRow(int n) {
vector<int> v;
for (int i = 0; i < n; ++i)
v.push_back(comb(n,i));
return v;
}
};
faster than 100% submissions on leet code https://leetcode.com/submissions/detail/406399031/
The most efficient way to calculate a row in pascal's triangle is through convolution. First we chose the second row (1,1) to be a kernel and then in order to get the next row we only need to convolve curent row with the kernel.
So convolution of the kernel with second row gives third row [1 1]*[1 1] = [1 2 1], convolution with the third row gives fourth [1 2 1]*[1 1] = [1 3 3 1] and so on
This is a function in julia-lang (very simular to matlab):
function binomRow(n::Int64)
baseVector = [1] #the first row is equal to 1.
kernel = [1,1] #This is the second row and a kernel.
row = zeros(n)
for i = 1 : n
row = baseVector
baseVector = conv(baseVector, kernel) #convoltion with kernel
end
return row::Array{Int64,1}
end
To find nth row -
int res[] = new int[n+1];
res[0] = 1;
for(int i = 1; i <= n; i++)
for(int j = i; j > 0; j++)
res[j] += res[j-1];
I need to be able to process one word or words and verify that it has valid syllables. There are some syllabification rules that could be used:
V CV VC CVC CCV CCCV CVCC
where V is a vowel and C is a consonant. e.g.,
pronunciation (5 Pro-nun-ci-a-tion; CCV-CVC-CV-V-CVC)
Or is there a simple code that can be used, or a library in c++? In class we're talking about binary search trees, hash tables, etc, but i can't really see the relation. Any help would appreciated, thanks.
Whenever we have collected a full pattern-string, we can either discard it and begin collecting to a new pattern-string, or keep it and try to get a longer pattern-string. We don't know in advance (without examining the rest of the input-string), whether we should keep or discard the current string, so we need to keep both possibilities in mind.
We can build a state machine that can keep track of this for us. The base-states are identified by the sequence of characters we have examined so far:
State C V
"" {"C"} {"V",""}
"C" {"CC"} {"CV",""}
"CC" {"CCC"} {""}
"CCC" {} {""}
"CV" {"CVC",""} {}
"CVC" {""} {}
"V" {""} {}
Since we don't always know which action to take, we can be in several possible states at once. Those sets of possible states form super-states:
Index Super-state C V
0 {} 0 0 Fail
1 {""} 2 9 Accept
2 {"C"} 3 8
3 {"CC"} 4 1
4 {"CCC"} 0 1
5 {"","C"} 6 13 Accept
6 {"C","CC"} 7 8
7 {"CC","CCC"} 4 1
8 {"","CV"} 12 9 Accept
9 {"","V"} 5 9 Accept
10 {"","C","CC"} 11 13 Accept
11 {"C","CC","CCC"} 7 8
12 {"","C","CVC"} 10 13 Accept
13 {"","CV","V"} 12 9 Accept
The transitions are between super-states. Each member of the super-state is advanced with the same symbol. All members without such transition are discarded. If a member has two possible destinations, both are added to the new super-state.
You might notice that some rows are very similar. Super-state 3 and 7 have the same transitions. As are 6 and 11, and 8 and 13. You could collapse those into one state each, and update the indices. I'm not going to demonstrate that here.
This could easily be encoded into a programming language:
// index = 0 1 2 3 4 5 6 7 8 9 10 11 12 13
int[] consonant = new int[] { 0, 2, 3, 4, 0, 6, 7, 4, 12, 5, 11, 7, 10, 12 };
int[] vocal = new int[] { 0, 9, 8, 1, 1, 13, 8, 1, 9, 9, 13, 8, 13, 9 };
int[] accept = new int[] { 0, 1, 0, 0, 0, 1, 0, 0, 1, 1, 1, 0, 1, 1 };
int startState = 1;
int failState = 0;
bool CheckWord(string word)
{
int state = startState;
foreach (char c in word)
{
if (IsVocal(c))
{
state = vocal[state];
}
else if (IsConsonant(c))
{
state = consonant[state];
}
if (state == failState) return false;
}
return accept[state] != 0;
}
Example:
> CheckWord("pronunciation")
true
> CheckWord("pronunciationn")
false
I wrote Levenshtein algorithm in in C++
If I input:
string s: democrat
string t: republican
I get the matrix D filled-up and the number of operations (the Levenshtein distance) can be read in D[10][8] = 8
Beyond the filled matrix I want to construct the optimal solution. How must look this solution? I don't have an idea.
Please only write me HOW MUST LOOK for this example.
The question is
Given the matrix produced by the Levenshtein algorithm, how can one find "the optimal solution"?
i.e. how can we find the precise sequence of string operations: inserts, deletes and substitution [of a single letter], necessary to convert the 's string' into the 't string'?
First, it should be noted that in many cases there are SEVERAL optimal solutions. While the Levenshtein algorithm supplies the minimum number of operations (8 in democrat/republican example) there are many sequences (of 8 operations) which can produce this conversion.
By "decoding" the Levenshtein matrix, one can enumerate ALL such optimal sequences.
The general idea is that the optimal solutions all follow a "path", from top left corner to bottom right corner (or in the other direction), whereby the matrix cell values on this path either remain the same or increase by one (or decrease by one in the reverse direction), starting at 0 and ending at the optimal number of operations for the strings in question (0 thru 8 democrat/republican case). The number increases when an operation is necessary, it stays the same when the letter at corresponding positions in the strings are the same.
It is easy to produce an algorithm which produces such a path (slightly more complicated to produce all possible paths), and from such path deduce the sequence of operations.
This path finding algorithm should start at the lower right corner and work its way backward. The reason for this approach is that we know for a fact that to be an optimal solution it must end in this corner, and to end in this corner, it must have come from one of the 3 cells either immediately to its left, immediately above it or immediately diagonally. By selecting a cell among these three cells, one which satisfies our "same value or decreasing by one" requirement, we effectively pick a cell on one of the optimal paths. By repeating the operation till we get on upper left corner (or indeed until we reach a cell with a 0 value), we effectively backtrack our way on an optimal path.
Illustration with the democrat - republican example
It should also be noted that one can build the matrix in one of two ways: with 'democrat' horizontally or vertically. This doesn't change the computation of the Levenshtein distance nor does it change the list of operations needed; it only changes the way we interpret the matrix, for example moving horizontally on the "path" either means inserting a character [from the t string] or deleting a character [off the s string] depending whether 'string s' is "horizontal" or "vertical" in the matrix.
I'll use the following matrix. The conventions are therefore (only going in the left-to-right and/or top-to-bottom directions)
an horizontal move is an INSERTION of a letter from the 't string'
an vertical move is a DELETION of a letter from the 's string'
a diagonal move is either:
a no-operation (both letters at respective positions are the same); the number doesn't change
a SUBSTITUTION (letters at respective positions are distinct); the number increase by one.
Levenshtein matrix for s = "democrat", t="republican"
r e p u b l i c a n
0 1 2 3 4 5 6 7 8 9 10
d 1 1 2 3 4 5 6 7 8 9 10
e 2 2 1 2 3 4 5 6 7 8 9
m 3 3 2 2 3 4 5 6 7 8 9
o 4 4 3 3 3 4 5 6 7 8 9
c 5 5 4 4 4 4 5 6 6 7 8
r 6 5 5 5 5 5 5 6 7 7 8
a 7 6 6 6 6 6 6 6 7 7 8
t 8 7 7 7 7 7 7 7 7 8 8
The arbitrary approach I use to select one path among several possible optimal paths is loosely described below:
Starting at the bottom-rightmost cell, and working our way backward toward
the top left.
For each "backward" step, consider the 3 cells directly adjacent to the current
cell (in the left, top or left+top directions)
if the value in the diagonal cell (going up+left) is smaller or equal to the
values found in the other two cells
AND
if this is same or 1 minus the value of the current cell
then "take the diagonal cell"
if the value of the diagonal cell is one less than the current cell:
Add a SUBSTITUTION operation (from the letters corresponding to
the _current_ cell)
otherwise: do not add an operation this was a no-operation.
elseif the value in the cell to the left is smaller of equal to the value of
the of the cell above current cell
AND
if this value is same or 1 minus the value of the current cell
then "take the cell to left", and
add an INSERTION of the letter corresponding to the cell
else
take the cell above, add
Add a DELETION operation of the letter in 's string'
Following this informal pseudo-code, we get the following:
Start on the "n", "t" cell at bottom right.
Pick the [diagonal] "a", "a" cell as next destination since it is less than the other two (and satisfies the same or -1 condition).
Note that the new cell is one less than current cell
therefore the step 8 is substitute "t" with "n": democra N
Continue with "a", "a" cell,
Pick the [diagonal] "c", "r" cell as next destination...
Note that the new cell is same value as current cell ==> no operation needed.
Continue with "c", "r" cell,
Pick the [diagonal] "i", "c" cell as next destination...
Note that the new cell is one less than current cell
therefore the step 7 is substitute "r" with "c": democ C an
Continue with "i", "c" cell,
Pick the [diagonal] "l", "o" cell as next destination...
Note that the new cell is one less than current cell
therefore the step 6 is substitute "c" with "i": demo I can
Continue with "l", "o" cell,
Pick the [diagonal] "b", "m" cell as next destination...
Note that the new cell is one less than current cell
therefore the step 5 is substitute "o" with "l": dem L ican
Continue with "b", "m" cell,
Pick the [diagonal]"u", "e" cell as next destination...
Note that the new cell is one less than current cell
therefore the step 4 is substitute "m" with "b": de B lican
Continue with "u", "e" cell,
Note the "diagonal" cell doesn't qualify, because the "left" cell is less than it.
Pick the [left] "p", "e" cell as next destination...
therefore the step 3 is instert "u" after "e": de U blican
Continue with "p", "e" cell,
again the "diagonal" cell doesn't qualify
Pick the [left] "e", "e" cell as next destination...
therefore the step 2 is instert "p" after "e": de P ublican
Continue with "e", "e" cell,
Pick the [diagonal] "r", "d" cell as next destination...
Note that the new cell is same value as current cell ==> no operation needed.
Continue with "r", "d" cell,
Pick the [diagonal] "start" cell as next destination...
Note that the new cell is one less than current cell
therefore the step 1 is substitute "d" with "r": R epublican
You've arrived at a cell which value is 0 : your work is done!
The backtracking algorithm to infer the moves from the matrix implemented in python:
def _backtrack_string(matrix, output_word):
'''
Iteratively backtrack DP matrix to get optimal set of moves
Inputs: DP matrix (list:list:int),
Input word (str),
Output word (str),
Start x position in DP matrix (int),
Start y position in DP matrix (int)
Output: Optimal path (list)
'''
i = len(matrix) - 1
j = len(matrix[0]) - 1
optimal_path = []
while i > 0 and j > 0:
diagonal = matrix[i-1][j-1]
vertical = matrix[i-1][j]
horizontal = matrix[i][j-1]
current = matrix[i][j]
if diagonal <= vertical and diagonal <= horizontal and (diagonal <= current):
i = i - 1
j = j - 1
if diagonal == current - 1:
optimal_path.append("Replace " + str(j) + ", " + str(output_word[j]) )
elif horizontal <= vertical and horizontal <= current:
j = j - 1
optimal_path.append("Insert " + str(j) + ", " + str(output_word[j]))
elif vertical <= horizontal and vertical <= current:
i = i - 1
optimal_path.append("Delete " + str(i))
elif horizontal <= vertical and horizontal <= current:
j = j - 1
optimal_path.append("Insert " + str(j) + ", " + str(output_word[j]))
else:
i = i - 1
optimal_path.append("Delete " + str(i))
return reversed(optimal_path)
The output I get when I run the algorithm with original word "OPERATING" and desired word "CONSTANTINE" is the following
Insert 0, C
Replace 2, N
Replace 3, S
Replace 4, T
Insert 6, N
Replace 10, E
"" C O N S T A N T I N E
"" [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11]
<-- Insert 0, C
O [1, 1, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10]
\ Replace 2, N
P [2, 2, 2, 2, 3, 4, 5, 6, 7, 8, 9, 10]
\ Replace 3, S
E [3, 3, 3, 3, 3, 4, 5, 6, 7, 8, 9, 9]
\ Replace 4, T
R [4, 4, 4, 4, 4, 4, 5, 6, 7, 8, 9, 10] No move
\ <-- Insert 6, N
A [5, 5, 5, 5, 5, 5, 4, 5, 6, 7, 8, 9]
\ No move
T [6, 6, 6, 6, 6, 5, 5, 5, 5, 6, 7, 8]
\ No move
I [7, 7, 7, 7, 7, 6, 6, 6, 6, 5, 6, 7]
\ No move
N [8, 8, 8, 7, 8, 7, 7, 6, 7, 6, 5, 6]
\ Replace 10, E
G [9, 9, 9, 8, 8, 8, 8, 7, 7, 7, 6, 6]
Note that I had to add extra conditions if the element in the diagonal is the same as the current element. There could be a deletion or insertion depending on values in the vertical (up) and horizontal (left) positions. We only get a "no operation" or "replace" operation when the following occurs
# assume bottom right of a 2x2 matrix is the reference position
# and has value v
# the following is the situation where we get a replace operation
[v + 1 , v<]
[ v< , v]
# the following is the situation where we get a "no operation"
[v , v<]
[v<, v ]
I think this is where the algorithm described in the first answer could break. There could be other arrangements in the 2x2 matrix above when neither operations are correct. The example shown with input "OPERATING" and output "CONSTANTINE" breaks the algorithm unless this is taken into account.
It's been some times since I played with it, but it seems to me the matrix should look something like:
. . r e p u b l i c a n
. 0 1 2 3 4 5 6 7 8 9 10
d 1 1 2 3 4 5 6 7 8 9 10
e 2 2 1 2 3 4 5 6 7 8 9
m 3 3 2 2 3 4 5 6 7 8 9
o 4 4 3 3 3 4 5 6 7 8 9
c 5 5 4 4 4 4 5 6 7 8 9
r 6 5 5 5 5 5 5 6 7 8 9
a 7 6 6 6 6 6 6 6 7 7 8
t 8 7 7 7 7 7 7 7 7 7 8
Don't take it for granted though.
Here is a VBA algorithm based on mjv's answer.
(very well explained, but some case were missing).
Sub TU_Levenshtein()
Call Levenshtein("democrat", "republican")
Call Levenshtein("ooo", "u")
Call Levenshtein("ceci est un test", "ceci n'est pas un test")
End Sub
Sub Levenshtein(ByVal string1 As String, ByVal string2 As String)
' Fill Matrix Levenshtein (-> array 'Distance')
Dim i As Long, j As Long
Dim string1_length As Long
Dim string2_length As Long
Dim distance() As Long
string1_length = Len(string1)
string2_length = Len(string2)
ReDim distance(string1_length, string2_length)
For i = 0 To string1_length
distance(i, 0) = i
Next
For j = 0 To string2_length
distance(0, j) = j
Next
For i = 1 To string1_length
For j = 1 To string2_length
If Asc(Mid$(string1, i, 1)) = Asc(Mid$(string2, j, 1)) Then
distance(i, j) = distance(i - 1, j - 1)
Else
distance(i, j) = Application.WorksheetFunction.min _
(distance(i - 1, j) + 1, _
distance(i, j - 1) + 1, _
distance(i - 1, j - 1) + 1)
End If
Next
Next
LevenshteinDistance = distance(string1_length, string2_length) ' for information only
' Write Matrix on VBA sheets (only for visuation, not used in calculus)
Cells.Clear
For i = 1 To UBound(distance, 1)
Cells(i + 2, 1).Value = Mid(string1, i, 1)
Next i
For i = 1 To UBound(distance, 2)
Cells(1, i + 2).Value = Mid(string2, i, 1)
Next i
For i = 0 To UBound(distance, 1)
For j = 0 To UBound(distance, 2)
Cells(i + 2, j + 2) = distance(i, j)
Next j
Next i
' One solution
current_posx = UBound(distance, 1)
current_posy = UBound(distance, 2)
Do
cc = distance(current_posx, current_posy)
Cells(current_posx + 1, current_posy + 1).Interior.Color = vbYellow ' visualisation again
' Manage border case
If current_posy - 1 < 0 Then
MsgBox ("deletion. " & Mid(string1, current_posx, 1))
current_posx = current_posx - 1
current_posy = current_posy
GoTo suivant
End If
If current_posx - 1 < 0 Then
MsgBox ("insertion. " & Mid(string2, current_posy, 1))
current_posx = current_posx
current_posy = current_posy - 1
GoTo suivant
End If
' Middle cases
cc_L = distance(current_posx, current_posy - 1)
cc_U = distance(current_posx - 1, current_posy)
cc_D = distance(current_posx - 1, current_posy - 1)
If (cc_D <= cc_L And cc_D <= cc_U) And (cc_D = cc - 1 Or cc_D = cc) Then
If (cc_D = cc - 1) Then
MsgBox "substitution. " & Mid(string1, current_posx, 1) & " by " & Mid(string2, current_posy, 1)
current_posx = current_posx - 1
current_posy = current_posy - 1
GoTo suivant
Else
MsgBox "no operation"
current_posx = current_posx - 1
current_posy = current_posy - 1
GoTo suivant
End If
ElseIf cc_L <= cc_D And cc_L = cc - 1 Then
MsgBox ("insertion. " & Mid(string2, current_posy, 1))
current_posx = current_posx
current_posy = current_posy - 1
GoTo suivant
Else
MsgBox ("deletion." & Mid(string1, current_posy, 1))
current_posx = current_posx
current_posy = current_posy - 1
GoTo suivant
End If
suivant:
Loop While Not (current_posx = 0 And current_posy = 0)
End Sub
I've done some work with the Levenshtein distance algorithm's matrix recently. I needed to produce the operations which would transform one list into another. (This will work for strings too.)
Do the following (vows) tests show the sort of functionality that you're looking for?
, "lev - complex 2"
: { topic
: lev.diff([13, 6, 5, 1, 8, 9, 2, 15, 12, 7, 11], [9, 13, 6, 5, 1, 8, 2, 15, 12, 11])
, "check actions"
: function(topic) { assert.deepEqual(topic, [{ op: 'delete', pos: 9, val: 7 },
{ op: 'delete', pos: 5, val: 9 },
{ op: 'insert', pos: 0, val: 9 },
]); }
}
, "lev - complex 3"
: { topic
: lev.diff([9, 13, 6, 5, 1, 8, 2, 15, 12, 11], [13, 6, 5, 1, 8, 9, 2, 15, 12, 7, 11])
, "check actions"
: function(topic) { assert.deepEqual(topic, [{ op: 'delete', pos: 0, val: 9 },
{ op: 'insert', pos: 5, val: 9 },
{ op: 'insert', pos: 9, val: 7 }
]); }
}
, "lev - complex 4"
: { topic
: lev.diff([9, 13, 6, 5, 1, 8, 2, 15, 12, 11, 16], [13, 6, 5, 1, 8, 9, 2, 15, 12, 7, 11, 17])
, "check actions"
: function(topic) { assert.deepEqual(topic, [{ op: 'delete', pos: 0, val: 9 },
{ op: 'insert', pos: 5, val: 9 },
{ op: 'insert', pos: 9, val: 7 },
{ op: 'replace', pos: 11, val: 17 }
]); }
}
Here is some Matlab code, is this correct by your opinion? Seems to give the right results :)
clear all
s = char('democrat');
t = char('republican');
% Edit Matrix
m=length(s);
n=length(t);
mat=zeros(m+1,n+1);
for i=1:1:m
mat(i+1,1)=i;
end
for j=1:1:n
mat(1,j+1)=j;
end
for i=1:m
for j=1:n
if (s(i) == t(j))
mat(i+1,j+1)=mat(i,j);
else
mat(i+1,j+1)=1+min(min(mat(i+1,j),mat(i,j+1)),mat(i,j));
end
end
end
% Edit Sequence
s = char('democrat');
t = char('republican');
i = m+1;
j = n+1;
display([s ' --> ' t])
while(i ~= 1 && j ~= 1)
temp = min(min(mat(i-1,j-1), mat(i,j-1)), mat(i-1,j));
if(mat(i-1,j) == temp)
i = i - 1;
t = [t(1:j-1) s(i) t(j:end)];
disp(strcat(['iinsertion: i=' int2str(i) ' , j=' int2str(j) ' ; ' s ' --> ' t]))
elseif(mat(i-1,j-1) == temp)
if(mat(i-1,j-1) == mat(i,j))
i = i - 1;
j = j - 1;
disp(strcat(['uunchanged: i=' int2str(i) ' , j=' int2str(j) ' ; ' s ' --> ' t]))
else
i = i - 1;
j = j - 1;
t(j) = s(i);
disp(strcat(['substition: i=' int2str(i) ' , j=' int2str(j) ' ; ' s ' --> ' t]))
end
elseif(mat(i,j-1) == temp)
j = j - 1;
t(j) = [];
disp(strcat(['dddeletion: i=' int2str(i) ' , j=' int2str(j) ' ; ' s ' --> ' t]))
end
end
C# implementation of JackIsJack answer with some changes:
Operations are output in 'forward' order (JackIsJack outputs in reverse order);
Last 'else' clause in original answer worked incorrectly (looks like copy-paste error).
Console application code:
class Program
{
static void Main(string[] args)
{
Levenshtein("1", "1234567890");
Levenshtein( "1234567890", "1");
Levenshtein("kitten", "mittens");
Levenshtein("mittens", "kitten");
Levenshtein("kitten", "sitting");
Levenshtein("sitting", "kitten");
Levenshtein("1234567890", "12356790");
Levenshtein("12356790", "1234567890");
Levenshtein("ceci est un test", "ceci n'est pas un test");
Levenshtein("ceci n'est pas un test", "ceci est un test");
}
static void Levenshtein(string string1, string string2)
{
Console.WriteLine("Levenstein '" + string1 + "' => '" + string2 + "'");
var string1_length = string1.Length;
var string2_length = string2.Length;
int[,] distance = new int[string1_length + 1, string2_length + 1];
for (int i = 0; i <= string1_length; i++)
{
distance[i, 0] = i;
}
for (int j = 0; j <= string2_length; j++)
{
distance[0, j] = j;
}
for (int i = 1; i <= string1_length; i++)
{
for (int j = 1; j <= string2_length; j++)
{
if (string1[i - 1] == string2[j - 1])
{
distance[i, j] = distance[i - 1, j - 1];
}
else
{
distance[i, j] = Math.Min(distance[i - 1, j] + 1, Math.Min(
distance[i, j - 1] + 1,
distance[i - 1, j - 1] + 1));
}
}
}
var LevenshteinDistance = distance[string1_length, string2_length];// for information only
Console.WriteLine($"Levernstein distance: {LevenshteinDistance}");
// List of operations
var current_posx = string1_length;
var current_posy = string2_length;
var stack = new Stack<string>(); // for outputting messages in forward direction
while (current_posx != 0 || current_posy != 0)
{
var cc = distance[current_posx, current_posy];
// edge cases
if (current_posy - 1 < 0)
{
stack.Push("Delete '" + string1[current_posx - 1] + "'");
current_posx--;
continue;
}
if (current_posx - 1 < 0)
{
stack.Push("Insert '" + string2[current_posy - 1] + "'");
current_posy--;
continue;
}
// Middle cases
var cc_L = distance[current_posx, current_posy - 1];
var cc_U = distance[current_posx - 1, current_posy];
var cc_D = distance[current_posx - 1, current_posy - 1];
if ((cc_D <= cc_L && cc_D <= cc_U) && (cc_D == cc - 1 || cc_D == cc))
{
if (cc_D == cc - 1)
{
stack.Push("Substitute '" + string1[current_posx - 1] + "' by '" + string2[current_posy - 1] + "'");
current_posx--;
current_posy--;
}
else
{
stack.Push("Keep '" + string1[current_posx - 1] + "'");
current_posx--;
current_posy--;
}
}
else if (cc_L <= cc_D && cc_L == cc - 1)
{
stack.Push("Insert '" + string2[current_posy - 1] + "'");
current_posy--;
}
else
{
stack.Push("Delete '" + string1[current_posx - 1]+"'");
current_posx--;
}
}
while(stack.Count > 0)
{
Console.WriteLine(stack.Pop());
}
}
}
The code to get all the edit paths according to edit matrix, source and target. Make a comment if there are any bugs. Thanks a lot!
import copy
from typing import List, Union
def edit_distance(source: Union[List[str], str],
target: Union[List[str], str],
return_distance: bool = False):
"""get the edit matrix
"""
edit_matrix = [[i + j for j in range(len(target) + 1)] for i in range(len(source) + 1)]
for i in range(1, len(source) + 1):
for j in range(1, len(target) + 1):
if source[i - 1] == target[j - 1]:
d = 0
else:
d = 1
edit_matrix[i][j] = min(edit_matrix[i - 1][j] + 1,
edit_matrix[i][j - 1] + 1,
edit_matrix[i - 1][j - 1] + d)
if return_distance:
return edit_matrix[len(source)][len(target)]
return edit_matrix
def get_edit_paths(matrix: List[List[int]],
source: Union[List[str], str],
target: Union[List[str], str]):
"""get all the valid edit paths
"""
all_paths = []
def _edit_path(i, j, optimal_path):
if i > 0 and j > 0:
diagonal = matrix[i - 1][j - 1] # the diagonal value
vertical = matrix[i - 1][j] # the above value
horizontal = matrix[i][j - 1] # the left value
current = matrix[i][j] # current value
# whether the source and target token are the same
flag = False
# compute the minimal value of the diagonal, vertical and horizontal
minimal = min(diagonal, min(vertical, horizontal))
# if the diagonal is the minimal
if diagonal == minimal:
new_i = i - 1
new_j = j - 1
path_ = copy.deepcopy(optimal_path)
# if the diagnoal value equals to current - 1
# it means `replace`` operation
if diagonal == current - 1:
path_.append(f"Replace | {new_j} | {target[new_j]}")
_edit_path(new_i, new_j, path_)
# if the diagonal value equals to current value
# and corresponding positional value of source and target equal
# it means this is current best path
elif source[new_i] == target[new_j]:
flag = True
# path_.append(f"Keep | {new_i}")
_edit_path(new_i, new_j, path_)
# if the position doesn't have best path
# we need to consider other situations
if not flag:
# if vertical value equals to minimal
# it means delete source corresponding value
if vertical == minimal:
new_i = i - 1
new_j = j
path_ = copy.deepcopy(optimal_path)
path_.append(f"Delete | {new_i}")
_edit_path(new_i, new_j, path_)
# if horizontal value equals to minimal
# if mean insert target corresponding value to source
if horizontal == minimal:
new_i = i
new_j = j - 1
path_ = copy.deepcopy(optimal_path)
path_.append(f"Insert | {new_j} | {target[new_j]}")
_edit_path(new_i, new_j, path_)
else:
all_paths.append(list(reversed(optimal_path)))
# get the rows and columns of the edit matrix
row_len = len(matrix) - 1
col_len = len(matrix[0]) - 1
_edit_path(row_len, col_len, optimal_path=[])
return all_paths
if __name__ == "__main__":
source = "BBDEF"
target = "ABCDF"
matrix = edit_distance(source, target)
print("print paths")
paths = get_edit_paths(matrix, source=list(source), target=list(target))
for path in paths:
print(path)