Having some difficulty troubleshooting code I wrote in C to perform a logistic regression. While it seems to work on smaller, semi-randomized datasets, it stops working (e.g. assigning proper probabilities of belonging to class 1) at around the point where I pass 43,500 observations (determined by tweaking the number of observations created. When creating the 150 features used in the code, I do create the first two as a function of the number of observations, so I'm not sure if maybe that's the issue here, though I am using double precision. Maybe there's an overflow somewhere in the code?
The below code should be self-contained; it generates m=50,000 observations with n=150 features. Setting m below 43,500 should return "Percent class 1: 0.250000", setting to 44,000 or above will return "Percent class 1: 0.000000", regardless of what max_iter (number of times we sample m observations) is set to.
The first feature is set to 1.0 divided by the total number of observations, if class 0 (first 75% of observations), or the index of the observation divided by the total number of observations otherwise.
The second feature is just index divided by total number of observations.
All other features are random.
The logistic regression is intended to use stochastic gradient descent, randomly selecting an observation index, computing the gradient of the loss with the predicted y using current weights, and updating weights with the gradient and learning rate (eta).
Using the same initialization with Python and NumPy, I still get the proper results, even above 50,000 observations.
#include <stdlib.h>
#include <stdio.h>
#include <math.h>
#include <time.h>
// Compute z = w * x + b
double dlc( int n, double *X, double *coef, double intercept )
{
double y_pred = intercept;
for (int i = 0; i < n; i++)
{
y_pred += X[i] * coef[i];
}
return y_pred;
}
// Compute y_hat = 1 / (1 + e^(-z))
double sigmoid( int n, double alpha, double *X, double *coef, double beta, double intercept )
{
double y_pred;
y_pred = dlc(n, X, coef, intercept);
y_pred = 1.0 / (1.0 + exp(-y_pred));
return y_pred;
}
// Stochastic gradient descent
void sgd( int m, int n, double *X, double *y, double *coef, double *intercept, double eta, int max_iter, int fit_intercept, int random_seed )
{
double *gradient_coef, *X_i;
double y_i, y_pred, resid;
int idx;
double gradient_intercept = 0.0, alpha = 1.0, beta = 1.0;
X_i = (double *) malloc (n * sizeof(double));
gradient_coef = (double *) malloc (n * sizeof(double));
for ( int i = 0; i < n; i++ )
{
coef[i] = 0.0;
gradient_coef[i] = 0.0;
}
*intercept = 0.0;
srand(random_seed);
for ( int epoch = 0; epoch < max_iter; epoch++ )
{
for ( int run = 0; run < m; run++ )
{
// Randomly sample an observation
idx = rand() % m;
for ( int i = 0; i < n; i++ )
{
X_i[i] = X[n*idx+i];
}
y_i = y[idx];
// Compute y_hat
y_pred = sigmoid( n, alpha, X_i, coef, beta, *intercept );
resid = -(y_i - y_pred);
// Compute gradients and adjust weights
for (int i = 0; i < n; i++)
{
gradient_coef[i] = X_i[i] * resid;
coef[i] -= eta * gradient_coef[i];
}
if ( fit_intercept == 1 )
{
*intercept -= eta * resid;
}
}
}
}
int main(void)
{
double *X, *y, *coef, *y_pred;
double intercept;
double eta = 0.05;
double alpha = 1.0, beta = 1.0;
long m = 50000;
long n = 150;
int max_iter = 20;
long class_0 = (long)(3.0 / 4.0 * (double)m);
double pct_class_1 = 0.0;
clock_t test_start;
clock_t test_end;
double test_time;
printf("Constructing variables...\n");
X = (double *) malloc (m * n * sizeof(double));
y = (double *) malloc (m * sizeof(double));
y_pred = (double *) malloc (m * sizeof(double));
coef = (double *) malloc (n * sizeof(double));
// Initialize classes
for (int i = 0; i < m; i++)
{
if (i < class_0)
{
y[i] = 0.0;
}
else
{
y[i] = 1.0;
}
}
// Initialize observation features
for (int i = 0; i < m; i++)
{
if (i < class_0)
{
X[n*i] = 1.0 / (double)m;
}
else
{
X[n*i] = (double)i / (double)m;
}
X[n*i + 1] = (double)i / (double)m;
for (int j = 2; j < n; j++)
{
X[n*i + j] = (double)(rand() % 100) / 100.0;
}
}
// Fit weights
printf("Running SGD...\n");
test_start = clock();
sgd( m, n, X, y, coef, &intercept, eta, max_iter, 1, 42 );
test_end = clock();
test_time = (double)(test_end - test_start) / CLOCKS_PER_SEC;
printf("Time taken: %f\n", test_time);
// Compute y_hat and share of observations predicted as class 1
printf("Making predictions...\n");
for ( int i = 0; i < m; i++ )
{
y_pred[i] = sigmoid( n, alpha, &X[i*n], coef, beta, intercept );
}
printf("Printing results...\n");
for ( int i = 0; i < m; i++ )
{
//printf("%f\n", y_pred[i]);
if (y_pred[i] > 0.5)
{
pct_class_1 += 1.0;
}
// Troubleshooting print
if (i < 10 || i > m - 10)
{
printf("%g\n", y_pred[i]);
}
}
printf("Percent class 1: %f", pct_class_1 / (double)m);
return 0;
}
For reference, here is my (presumably) equivalent Python code, which returns the correct percent of identified classes at more than 50,000 observations:
import numpy as np
import time
def sigmoid(x):
return 1 / (1 + np.exp(-x))
class LogisticRegressor:
def __init__(self, eta, init_runs, fit_intercept=True):
self.eta = eta
self.init_runs = init_runs
self.fit_intercept = fit_intercept
def fit(self, x, y):
m, n = x.shape
self.coef = np.zeros((n, 1))
self.intercept = np.zeros((1, 1))
for epoch in range(self.init_runs):
for run in range(m):
idx = np.random.randint(0, m)
x_i = x[idx:idx+1, :]
y_i = y[idx]
y_pred_i = sigmoid(x_i.dot(self.coef) + self.intercept)
gradient_w = -(x_i.T * (y_i - y_pred_i))
self.coef -= self.eta * gradient_w
if self.fit_intercept:
gradient_b = -(y_i - y_pred_i)
self.intercept -= self.eta * gradient_b
def predict_proba(self, x):
m, n = x.shape
y_pred = np.ones((m, 2))
y_pred[:,1:2] = sigmoid(x.dot(self.coef) + self.intercept)
y_pred[:,0:1] -= y_pred[:,1:2]
return y_pred
def predict(self, x):
return np.round(sigmoid(x.dot(self.coef) + self.intercept))
m = 50000
n = 150
class1 = int(3.0 / 4.0 * m)
X = np.random.rand(m, n)
y = np.zeros((m, 1))
for obs in range(m):
if obs < class1:
continue
else:
y[obs,0] = 1
for obs in range(m):
if obs < class1:
X[obs, 0] = 1.0 / float(m)
else:
X[obs, 0] = float(obs) / float(m)
X[obs, 1] = float(obs) / float(m)
logit = LogisticRegressor(0.05, 20)
start_time = time.time()
logit.fit(X, y)
end_time = time.time()
print(round(end_time - start_time, 2))
y_pred = logit.predict(X)
print("Percent:", y_pred.sum() / len(y_pred))
The issue is here:
// Randomly sample an observation
idx = rand() % m;
... in light of the fact that the OP's RAND_MAX is 32767. This is exacerbated by the fact that all of the class 0 observations are at the end.
All samples will be drawn from the first 32768 observations, and when the total number of observations is greater than that, the proportion of class 0 observations among those that can be sampled is less than 0.25. At 43691 total observations, there are no class 0 observations among those that can be sampled.
As a secondary issue, rand() % m does not yield a wholly uniform distribution if m does not evenly divide RAND_MAX + 1, though the effect of this issue will be much more subtle.
Bottom line: you need a better random number generator.
At minimum, you could consider combining the bits from two calls to rand() to yield an integer with sufficient range, but you might want to consider getting a third-party generator. There are several available.
Note: OP reports "m=50,000 observations with n=150 features.", so perhaps this is not the issue for OP, but I'll leave this answer up for reference when OP tries larger tasks.
A potential issue:
long overflow
m * n * sizeof(double) risks overflow when long is 32-bit and m*n > LONG_MAX (or about 46,341 if m, n are the same).
OP does report
A first step is to perform the multiplication using size_t math where we gain at least 1 more bit in the calculation.
// m * n * sizeof(double)
sizeof(double) * m * n
Yet unless OP's size_t is more than 32-bit, we still have trouble.
IAC, I recommend to use size_t for array sizing and indexing.
Check allocations for failure too.
Since RAND_MAX may be too small and array indexing should be done using size_t math, consider a helper function to generate a random index over the entire size_t range.
// idx = rand() % m;
size_t idx = rand_size_t() % (size_t)m;
If stuck with the standard rand(), below is a helper function to extend its range as needed.
It uses the real nifty IMAX_BITS(m).
#include <assert.h>
#include <limits.h>
#include <stdint.h>
#include <stdlib.h>
// https://stackoverflow.com/a/4589384/2410359
/* Number of bits in inttype_MAX, or in any (1<<k)-1 where 0 <= k < 2040 */
#define IMAX_BITS(m) ((m)/((m)%255+1) / 255%255*8 + 7-86/((m)%255+12))
// Test that RAND_MAX is a power of 2 minus 1
_Static_assert((RAND_MAX & 1) && ((RAND_MAX/2 + 1) & (RAND_MAX/2)) == 0, "RAND_MAX is not a Mersenne number");
#define RAND_MAX_WIDTH (IMAX_BITS(RAND_MAX))
#define SIZE_MAX_WIDTH (IMAX_BITS(SIZE_MAX))
size_t rand_size_t(void) {
size_t index = (size_t) rand();
for (unsigned i = RAND_MAX_WIDTH; i < SIZE_MAX_WIDTH; i += RAND_MAX_WIDTH) {
index <<= RAND_MAX_WIDTH;
index ^= (size_t) rand();
}
return index;
}
Further considerations can replace the rand_size_t() % (size_t)m with a more uniform distribution.
As has been determined elsewhere, the problem is due to the implementation's RAND_MAX value being too small.
Assuming 32-bit ints, a slightly better PRNG function can be implemented in the code, such as this C implementation of the minstd_rand() function from C++:
#define MINSTD_RAND_MAX 2147483646
// Code assumes `int` is at least 32 bits wide.
static unsigned int minstd_seed = 1;
static void minstd_srand(unsigned int seed)
{
seed %= 2147483647;
// zero seed is bad!
minstd_seed = seed ? seed : 1;
}
static int minstd_rand(void)
{
minstd_seed = (unsigned long long)minstd_seed * 48271 % 2147483647;
return (int)minstd_seed;
}
Another problem is that expressions of the form rand() % m produce a biased result when m does not divide (unsigned int)RAND_MAX + 1. Here is an unbiased function that returns a random integer from 0 to le inclusive, making use of the minstd_rand() function defined earlier:
static int minstd_rand_max(int le)
{
int r;
if (le < 0)
{
r = le;
}
else if (le >= MINSTD_RAND_MAX)
{
r = minstd_rand();
}
else
{
int rm = MINSTD_RAND_MAX - le + MINSTD_RAND_MAX % (le + 1);
while ((r = minstd_rand()) > rm)
{
}
r /= (rm / (le + 1) + 1);
}
return r;
}
(Actually, it does still have a very small bias because minstd_rand() will never return 0.)
For example, replace rand() % 100 with minstd_rand_max(99), and replace rand() % m with minstd_rand_max(m - 1). Also replace srand(random_seed) with minstd_srand(random_seed).
I have a logical problem in my code, maybe it is caused by overflowing but I can't solve this on my own, so I would be thankful if anyone can help me.
In the following piece of code, I have implemented the function taylor_log(), which can count "n" iterations of taylor polynomial. In the void function I am looking for number of iterations (*limit) which is enough to count a logarithm with desired accuracy compared to log function from .
The thing is that sometimes UINT_MAX is not enough iterations to get the desired accuracy and at this point I want to let the user know that the number of needed iterations is higher than UINT_MAX. But my code don't work, for example for x = 1e+280, eps = 623. It just counts, counts and never give result.
TaylorPolynomial
double taylor_log(double x, unsigned int n){
double f_sum = 1.0;
double sum = 0.0;
for (unsigned int i = 1; i <= n; i++)
{
f_sum *= (x - 1) / x;
sum += f_sum / i;
}
return sum;
}
void guessIt(double x, double eps, unsigned int *limit){
*limit = 10;
double real_log = log(x);
double t_log = taylor_log(x, *limit);
while(myabs(real_log - t_log) > eps)
{
if (*limit == UINT_MAX)
{
*limit = 0;
break;
}
if (*limit >= UINT_MAX/2)
{
*limit = UINT_MAX;
t_log = taylor_log(x, *limit);
}
else
{
*limit = (*limit) *2;
t_log = taylor_log(x, *limit);
}
}
}
EDIT: Ok guys, thanks for your reactions so far. I have changed my code to this:
if (*limit == UINT_MAX-1)
{
*limit = 0;
break;
}
if (*limit >= UINT_MAX/2)
{
*limit = UINT_MAX-1;
t_log = taylor_log(x, *limit);
}
but it still doesn't work correctly, I have set printf to the beggining of taylor_log() function to see the value of "n" and its (..., 671088640, 1342177280, 2684354560, 5, 4, 3, 2, 2, 1, 2013265920, ...). Don't understand it..
This code below assigns the limit to UINT_MAX
if (*limit >= UINT_MAX/2)
{
*limit = UINT_MAX;
t_log = taylor_log(x, *limit);
}
And your for loop is defined like this:
for (unsigned int i = 1; i <= n; i++)
i will ALWAYS be less than or equal to UINT_MAX because there is never going to be a value of i that is greater than UINT_MAX. Because that's the largest value i could ever be. So there is certainly overflow and your loop exit condition is never met. i rolls over to zero and the process repeats indefinitely.
You should change your loop condition to i < n or change your limit to UINT_MAX - 1.
[Edit]
OP coded correctly but must insure a limited range (0.5 < x < 2.0 ?)
Below is a code version that self determines when to stop. Iteration count goes high near x near 0.5 and 2.0. The iteration count needed goes into the millions. Such the alternative coded far below.
double taylor_logA(double x) {
double f_sum = 1.0;
double sum = 0.0;
for (unsigned int i = 1; ; i++) {
f_sum *= (x - 1) / x;
double sum_before = sum;
sum += f_sum / i;
if (sum_before == sum) {
printf("%d\n", i);
break;
}
}
return sum;
}
Wrongalternative implementation of the series: Ref
Sample alternative - it converges faster.
double taylor_log2(double x, unsigned int n) {
double f_sum = 1.0;
double sum = 0.0;
for (unsigned int i = 1; i <= n; i++) {
f_sum *= (x - 1) / 1; // / 1 (or remove)
if (i & 1) sum += f_sum / i;
else sum -= f_sum / i; // subtract even terms
}
return sum;
}
A reasonable number of terms will converge as needed.
Alternatively, continue until terms are too small (maybe 50 or so)
double taylor_log3(double x) {
double f_sum = 1.0;
double sum = 0.0;
for (unsigned int i = 1; ; i++) {
double sum_before = sum;
f_sum *= x - 1;
if (i & 1) sum += f_sum / i;
else sum -= f_sum / i;
if (sum_before == sum) {
printf("%d\n", i);
break;
}
}
return sum;
}
Other improvements possible. example see More efficient series
First, using std::numeric_limits<unsigned int>::max() will make your code more c++-ish than c-ish. Second, you can use the integral type unsigned long long and std::numeric_limits<unsigned long long>::max() for the limit, which is pretty mush the limit for an integral type. If you want a higher limit, you may use long double. floating points also allows you to use infinity with std::numeric_limits<double>::infinity() note that infinity work with double, float and long double.
If neither of these types provide you the precision you need, look at boost::multiprecision
First of all, the Taylor series for the logarithm function only converges for values of 0 < x < 2, so it's quite possible that the eps precision is never hit.
Secondly, are you sure that it loops forever, instead of hitting the *limit >= UINT_MAX/2 after a very long time?
OP is using the series well outside its usable range of 0.5 x < 2.0 with calls like taylor_log(1e280, n)
Even within the range, x values near the limits of 0.5 and 2.0 converge very slowly needing millions+ of iterations. A precise log() will not result. Best to use the 2x range about 1.0.
Create a wrapper function to call the original function in its sweet range of sqrt(2)/2 < x < sqrt(2). Converges, worst case, with about 40 iterations.
#define SQRT_0_5 0.70710678118654752440084436210485
#define LN2 0.69314718055994530941723212145818
// Valid over the range (0...DBL_MAX]
double taylor_logB(double x, unsigned int n) {
int expo;
double signif = frexp(x, &expo);
if (signif < SQRT_0_5) {
signif *= 2;
expo--;
}
double y = taylor_log(signif,n);
y += expo*LN2;
return y;
}
I am trying to implement a linear least squares fit onto 2 arrays of data: time vs amplitude. The only technique I know so far is to test all of the possible m and b points in (y = m*x+b) and then find out which combination fits my data best so that it has the least error. However, I think iterating so many combinations is sometimes useless because it tests out everything. Are there any techniques to speed up the process that I don't know about? Thanks.
Try this code. It fits y = mx + b to your (x,y) data.
The arguments to linreg are
linreg(int n, REAL x[], REAL y[], REAL* b, REAL* m, REAL* r)
n = number of data points
x,y = arrays of data
*b = output intercept
*m = output slope
*r = output correlation coefficient (can be NULL if you don't want it)
The return value is 0 on success, !=0 on failure.
Here's the code
#include "linreg.h"
#include <stdlib.h>
#include <math.h> /* math functions */
//#define REAL float
#define REAL double
inline static REAL sqr(REAL x) {
return x*x;
}
int linreg(int n, const REAL x[], const REAL y[], REAL* m, REAL* b, REAL* r){
REAL sumx = 0.0; /* sum of x */
REAL sumx2 = 0.0; /* sum of x**2 */
REAL sumxy = 0.0; /* sum of x * y */
REAL sumy = 0.0; /* sum of y */
REAL sumy2 = 0.0; /* sum of y**2 */
for (int i=0;i<n;i++){
sumx += x[i];
sumx2 += sqr(x[i]);
sumxy += x[i] * y[i];
sumy += y[i];
sumy2 += sqr(y[i]);
}
REAL denom = (n * sumx2 - sqr(sumx));
if (denom == 0) {
// singular matrix. can't solve the problem.
*m = 0;
*b = 0;
if (r) *r = 0;
return 1;
}
*m = (n * sumxy - sumx * sumy) / denom;
*b = (sumy * sumx2 - sumx * sumxy) / denom;
if (r!=NULL) {
*r = (sumxy - sumx * sumy / n) / /* compute correlation coeff */
sqrt((sumx2 - sqr(sumx)/n) *
(sumy2 - sqr(sumy)/n));
}
return 0;
}
Example
You can run this example online.
int main()
{
int n = 6;
REAL x[6]= {1, 2, 4, 5, 10, 20};
REAL y[6]= {4, 6, 12, 15, 34, 68};
REAL m,b,r;
linreg(n,x,y,&m,&b,&r);
printf("m=%g b=%g r=%g\n",m,b,r);
return 0;
}
Here is the output
m=3.43651 b=-0.888889 r=0.999192
Here is the Excel plot and linear fit (for verification).
All values agree exactly with the C code above (note C code returns r while Excel returns R**2).
There are efficient algorithms for least-squares fitting; see Wikipedia for details. There are also libraries that implement the algorithms for you, likely more efficiently than a naive implementation would do; the GNU Scientific Library is one example, but there are others under more lenient licenses as well.
From Numerical Recipes: The Art of Scientific Computing in (15.2) Fitting Data to a Straight Line:
Linear Regression:
Consider the problem of fitting a set of N data points (xi, yi) to a straight-line model:
Assume that the uncertainty: sigmai associated with each yi and that the xi’s (values of the dependent variable) are known exactly. To measure how well the model agrees with the data, we use the chi-square function, which in this case is:
The above equation is minimized to determine a and b. This is done by finding the derivative of the above equation with respect to a and b, equate them to zero and solve for a and b. Then we estimate the probable uncertainties in the estimates of a and b, since obviously the measurement errors in the data must introduce some uncertainty in the determination of those parameters. Additionally, we must estimate the goodness-of-fit of the data to the
model. Absent this estimate, we have not the slightest indication that the parameters a and b in the model have any meaning at all.
The below struct performs the mentioned calculations:
struct Fitab {
// Object for fitting a straight line y = a + b*x to a set of
// points (xi, yi), with or without available
// errors sigma i . Call one of the two constructors to calculate the fit.
// The answers are then available as the variables:
// a, b, siga, sigb, chi2, and either q or sigdat.
int ndata;
double a, b, siga, sigb, chi2, q, sigdat; // Answers.
vector<double> &x, &y, &sig;
// Constructor.
Fitab(vector<double> &xx, vector<double> &yy, vector<double> &ssig)
: ndata(xx.size()), x(xx), y(yy), sig(ssig), chi2(0.), q(1.), sigdat(0.)
{
// Given a set of data points x[0..ndata-1], y[0..ndata-1]
// with individual standard deviations sig[0..ndata-1],
// sets a,b and their respective probable uncertainties
// siga and sigb, the chi-square: chi2, and the goodness-of-fit
// probability: q
Gamma gam;
int i;
double ss=0., sx=0., sy=0., st2=0., t, wt, sxoss; b=0.0;
for (i=0;i < ndata; i++) { // Accumulate sums ...
wt = 1.0 / SQR(sig[i]); //...with weights
ss += wt;
sx += x[i]*wt;
sy += y[i]*wt;
}
sxoss = sx/ss;
for (i=0; i < ndata; i++) {
t = (x[i]-sxoss) / sig[i];
st2 += t*t;
b += t*y[i]/sig[i];
}
b /= st2; // Solve for a, b, sigma-a, and simga-b.
a = (sy-sx*b) / ss;
siga = sqrt((1.0+sx*sx/(ss*st2))/ss);
sigb = sqrt(1.0/st2); // Calculate chi2.
for (i=0;i<ndata;i++) chi2 += SQR((y[i]-a-b*x[i])/sig[i]);
if (ndata>2) q=gam.gammq(0.5*(ndata-2),0.5*chi2); // goodness of fit
}
// Constructor.
Fitab(vector<double> &xx, vector<double> &yy)
: ndata(xx.size()), x(xx), y(yy), sig(xx), chi2(0.), q(1.), sigdat(0.)
{
// As above, but without known errors (sig is not used).
// The uncertainties siga and sigb are estimated by assuming
// equal errors for all points, and that a straight line is
// a good fit. q is returned as 1.0, the normalization of chi2
// is to unit standard deviation on all points, and sigdat
// is set to the estimated error of each point.
int i;
double ss,sx=0.,sy=0.,st2=0.,t,sxoss;
b=0.0; // Accumulate sums ...
for (i=0; i < ndata; i++) {
sx += x[i]; // ...without weights.
sy += y[i];
}
ss = ndata;
sxoss = sx/ss;
for (i=0;i < ndata; i++) {
t = x[i]-sxoss;
st2 += t*t;
b += t*y[i];
}
b /= st2; // Solve for a, b, sigma-a, and sigma-b.
a = (sy-sx*b)/ss;
siga=sqrt((1.0+sx*sx/(ss*st2))/ss);
sigb=sqrt(1.0/st2); // Calculate chi2.
for (i=0;i<ndata;i++) chi2 += SQR(y[i]-a-b*x[i]);
if (ndata > 2) sigdat=sqrt(chi2/(ndata-2));
// For unweighted data evaluate typical
// sig using chi2, and adjust
// the standard deviations.
siga *= sigdat;
sigb *= sigdat;
}
};
where struct Gamma:
struct Gamma : Gauleg18 {
// Object for incomplete gamma function.
// Gauleg18 provides coefficients for Gauss-Legendre quadrature.
static const Int ASWITCH=100; When to switch to quadrature method.
static const double EPS; // See end of struct for initializations.
static const double FPMIN;
double gln;
double gammp(const double a, const double x) {
// Returns the incomplete gamma function P(a,x)
if (x < 0.0 || a <= 0.0) throw("bad args in gammp");
if (x == 0.0) return 0.0;
else if ((Int)a >= ASWITCH) return gammpapprox(a,x,1); // Quadrature.
else if (x < a+1.0) return gser(a,x); // Use the series representation.
else return 1.0-gcf(a,x); // Use the continued fraction representation.
}
double gammq(const double a, const double x) {
// Returns the incomplete gamma function Q(a,x) = 1 - P(a,x)
if (x < 0.0 || a <= 0.0) throw("bad args in gammq");
if (x == 0.0) return 1.0;
else if ((Int)a >= ASWITCH) return gammpapprox(a,x,0); // Quadrature.
else if (x < a+1.0) return 1.0-gser(a,x); // Use the series representation.
else return gcf(a,x); // Use the continued fraction representation.
}
double gser(const Doub a, const Doub x) {
// Returns the incomplete gamma function P(a,x) evaluated by its series representation.
// Also sets ln (gamma) as gln. User should not call directly.
double sum,del,ap;
gln=gammln(a);
ap=a;
del=sum=1.0/a;
for (;;) {
++ap;
del *= x/ap;
sum += del;
if (fabs(del) < fabs(sum)*EPS) {
return sum*exp(-x+a*log(x)-gln);
}
}
}
double gcf(const Doub a, const Doub x) {
// Returns the incomplete gamma function Q(a, x) evaluated
// by its continued fraction representation.
// Also sets ln (gamma) as gln. User should not call directly.
int i;
double an,b,c,d,del,h;
gln=gammln(a);
b=x+1.0-a; // Set up for evaluating continued fraction
// by modified Lentz’s method with with b0 = 0.
c=1.0/FPMIN;
d=1.0/b;
h=d;
for (i=1;;i++) {
// Iterate to convergence.
an = -i*(i-a);
b += 2.0;
d=an*d+b;
if (fabs(d) < FPMIN) d=FPMIN;
c=b+an/c;
if (fabs(c) < FPMIN) c=FPMIN;
d=1.0/d;
del=d*c;
h *= del;
if (fabs(del-1.0) <= EPS) break;
}
return exp(-x+a*log(x)-gln)*h; Put factors in front.
}
double gammpapprox(double a, double x, int psig) {
// Incomplete gamma by quadrature. Returns P(a,x) or Q(a, x),
// when psig is 1 or 0, respectively. User should not call directly.
int j;
double xu,t,sum,ans;
double a1 = a-1.0, lna1 = log(a1), sqrta1 = sqrt(a1);
gln = gammln(a);
// Set how far to integrate into the tail:
if (x > a1) xu = MAX(a1 + 11.5*sqrta1, x + 6.0*sqrta1);
else xu = MAX(0.,MIN(a1 - 7.5*sqrta1, x - 5.0*sqrta1));
sum = 0;
for (j=0;j<ngau;j++) { // Gauss-Legendre.
t = x + (xu-x)*y[j];
sum += w[j]*exp(-(t-a1)+a1*(log(t)-lna1));
}
ans = sum*(xu-x)*exp(a1*(lna1-1.)-gln);
return (psig?(ans>0.0? 1.0-ans:-ans):(ans>=0.0? ans:1.0+ans));
}
double invgammp(Doub p, Doub a);
// Inverse function on x of P(a,x) .
};
const Doub Gamma::EPS = numeric_limits<Doub>::epsilon();
const Doub Gamma::FPMIN = numeric_limits<Doub>::min()/EPS
and stuct Gauleg18:
struct Gauleg18 {
// Abscissas and weights for Gauss-Legendre quadrature.
static const Int ngau = 18;
static const Doub y[18];
static const Doub w[18];
};
const Doub Gauleg18::y[18] = {0.0021695375159141994,
0.011413521097787704,0.027972308950302116,0.051727015600492421,
0.082502225484340941, 0.12007019910960293,0.16415283300752470,
0.21442376986779355, 0.27051082840644336, 0.33199876341447887,
0.39843234186401943, 0.46931971407375483, 0.54413605556657973,
0.62232745288031077, 0.70331500465597174, 0.78649910768313447,
0.87126389619061517, 0.95698180152629142};
const Doub Gauleg18::w[18] = {0.0055657196642445571,
0.012915947284065419,0.020181515297735382,0.027298621498568734,
0.034213810770299537,0.040875750923643261,0.047235083490265582,
0.053244713977759692,0.058860144245324798,0.064039797355015485
0.068745323835736408,0.072941885005653087,0.076598410645870640,
0.079687828912071670,0.082187266704339706,0.084078218979661945,
0.085346685739338721,0.085983275670394821};
and, finally fuinction Gamma::invgamp():
double Gamma::invgammp(double p, double a) {
// Returns x such that P(a,x) = p for an argument p between 0 and 1.
int j;
double x,err,t,u,pp,lna1,afac,a1=a-1;
const double EPS=1.e-8; // Accuracy is the square of EPS.
gln=gammln(a);
if (a <= 0.) throw("a must be pos in invgammap");
if (p >= 1.) return MAX(100.,a + 100.*sqrt(a));
if (p <= 0.) return 0.0;
if (a > 1.) {
lna1=log(a1);
afac = exp(a1*(lna1-1.)-gln);
pp = (p < 0.5)? p : 1. - p;
t = sqrt(-2.*log(pp));
x = (2.30753+t*0.27061)/(1.+t*(0.99229+t*0.04481)) - t;
if (p < 0.5) x = -x;
x = MAX(1.e-3,a*pow(1.-1./(9.*a)-x/(3.*sqrt(a)),3));
} else {
t = 1.0 - a*(0.253+a*0.12); and (6.2.9).
if (p < t) x = pow(p/t,1./a);
else x = 1.-log(1.-(p-t)/(1.-t));
}
for (j=0;j<12;j++) {
if (x <= 0.0) return 0.0; // x too small to compute accurately.
err = gammp(a,x) - p;
if (a > 1.) t = afac*exp(-(x-a1)+a1*(log(x)-lna1));
else t = exp(-x+a1*log(x)-gln);
u = err/t;
// Halley’s method.
x -= (t = u/(1.-0.5*MIN(1.,u*((a-1.)/x - 1))));
// Halve old value if x tries to go negative.
if (x <= 0.) x = 0.5*(x + t);
if (fabs(t) < EPS*x ) break;
}
return x;
}
Here is my version of a C/C++ function that does simple linear regression. The calculations follow the wikipedia article on simple linear regression. This is published as a single-header public-domain (MIT) library on github: simple_linear_regression. The library (.h file) is tested to work on Linux and Windows, and from C and C++ using -Wall -Werror and all -std versions supported by clang/gcc.
#define SIMPLE_LINEAR_REGRESSION_ERROR_INPUT_VALUE -2
#define SIMPLE_LINEAR_REGRESSION_ERROR_NUMERIC -3
int simple_linear_regression(const double * x, const double * y, const int n, double * slope_out, double * intercept_out, double * r2_out) {
double sum_x = 0.0;
double sum_xx = 0.0;
double sum_xy = 0.0;
double sum_y = 0.0;
double sum_yy = 0.0;
double n_real = (double)(n);
int i = 0;
double slope = 0.0;
double denominator = 0.0;
if (x == NULL || y == NULL || n < 2) {
return SIMPLE_LINEAR_REGRESSION_ERROR_INPUT_VALUE;
}
for (i = 0; i < n; ++i) {
sum_x += x[i];
sum_xx += x[i] * x[i];
sum_xy += x[i] * y[i];
sum_y += y[i];
sum_yy += y[i] * y[i];
}
denominator = n_real * sum_xx - sum_x * sum_x;
if (denominator == 0.0) {
return SIMPLE_LINEAR_REGRESSION_ERROR_NUMERIC;
}
slope = (n_real * sum_xy - sum_x * sum_y) / denominator;
if (slope_out != NULL) {
*slope_out = slope;
}
if (intercept_out != NULL) {
*intercept_out = (sum_y - slope * sum_x) / n_real;
}
if (r2_out != NULL) {
denominator = ((n_real * sum_xx) - (sum_x * sum_x)) * ((n_real * sum_yy) - (sum_y * sum_y));
if (denominator == 0.0) {
return SIMPLE_LINEAR_REGRESSION_ERROR_NUMERIC;
}
*r2_out = ((n_real * sum_xy) - (sum_x * sum_y)) * ((n_real * sum_xy) - (sum_x * sum_y)) / denominator;
}
return 0;
}
Usage example:
#define SIMPLE_LINEAR_REGRESSION_IMPLEMENTATION
#include "simple_linear_regression.h"
#include <stdio.h>
/* Some data that we want to find the slope, intercept and r2 for */
static const double x[] = { 1.47, 1.50, 1.52, 1.55, 1.57, 1.60, 1.63, 1.65, 1.68, 1.70, 1.73, 1.75, 1.78, 1.80, 1.83 };
static const double y[] = { 52.21, 53.12, 54.48, 55.84, 57.20, 58.57, 59.93, 61.29, 63.11, 64.47, 66.28, 68.10, 69.92, 72.19, 74.46 };
int main() {
double slope = 0.0;
double intercept = 0.0;
double r2 = 0.0;
int res = 0;
res = simple_linear_regression(x, y, sizeof(x) / sizeof(x[0]), &slope, &intercept, &r2);
if (res < 0) {
printf("Error: %s\n", simple_linear_regression_error_string(res));
return res;
}
printf("slope: %f\n", slope);
printf("intercept: %f\n", intercept);
printf("r2: %f\n", r2);
return 0;
}
The original example above worked well for me with slope and offset but I had a hard time with the corr coef. Maybe I don't have my parenthesis working the same as the assumed precedence? Anyway, with some help of other web pages I finally got values that match the linear trend-line in Excel. Thought I would share my code using Mark Lakata's variable names. Hope this helps.
double slope = ((n * sumxy) - (sumx * sumy )) / denom;
double intercept = ((sumy * sumx2) - (sumx * sumxy)) / denom;
double term1 = ((n * sumxy) - (sumx * sumy));
double term2 = ((n * sumx2) - (sumx * sumx));
double term3 = ((n * sumy2) - (sumy * sumy));
double term23 = (term2 * term3);
double r2 = 1.0;
if (fabs(term23) > MIN_DOUBLE) // Define MIN_DOUBLE somewhere as 1e-9 or similar
r2 = (term1 * term1) / term23;
as an assignment I had to code in C a simple linear regression using RMSE loss function. The program is dynamic and you can enter your own values and choose your own loss function which is for now limited to Root Mean Square Error. But first here are the algorithms I used:
now the code... you need gnuplot to display the chart, sudo apt install gnuplot
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <math.h>
#include <sys/types.h>
#define BUFFSIZE 64
#define MAXSIZE 100
static double vector_x[MAXSIZE] = {0};
static double vector_y[MAXSIZE] = {0};
static double vector_predict[MAXSIZE] = {0};
static double max_x;
static double max_y;
static double mean_x;
static double mean_y;
static double teta_0_intercept;
static double teta_1_grad;
static double RMSE;
static double r_square;
static double prediction;
static char intercept[BUFFSIZE];
static char grad[BUFFSIZE];
static char xrange[BUFFSIZE];
static char yrange[BUFFSIZE];
static char lossname_RMSE[BUFFSIZE] = "Simple Linear Regression using RMSE'";
static char cmd_gnu_0[BUFFSIZE] = "set title '";
static char cmd_gnu_1[BUFFSIZE] = "intercept = ";
static char cmd_gnu_2[BUFFSIZE] = "grad = ";
static char cmd_gnu_3[BUFFSIZE] = "set xrange [0:";
static char cmd_gnu_4[BUFFSIZE] = "set yrange [0:";
static char cmd_gnu_5[BUFFSIZE] = "f(x) = (grad * x) + intercept";
static char cmd_gnu_6[BUFFSIZE] = "plot f(x), 'data.temp' with points pointtype 7";
static char const *commands_gnuplot[] = {
cmd_gnu_0,
cmd_gnu_1,
cmd_gnu_2,
cmd_gnu_3,
cmd_gnu_4,
cmd_gnu_5,
cmd_gnu_6,
};
static size_t size;
static void user_input()
{
printf("Enter x,y vector size, MAX = 100\n");
scanf("%lu", &size);
if (size > MAXSIZE) {
printf("Wrong input size is too big\n");
user_input();
}
printf("vector's size is %lu\n", size);
size_t i;
for (i = 0; i < size; i++) {
printf("Enter vector_x[%ld] values\n", i);
scanf("%lf", &vector_x[i]);
}
for (i = 0; i < size; i++) {
printf("Enter vector_y[%ld] values\n", i);
scanf("%lf", &vector_y[i]);
}
}
static void display_vector()
{
size_t i;
for (i = 0; i < size; i++){
printf("vector_x[%lu] = %lf\t", i, vector_x[i]);
printf("vector_y[%lu] = %lf\n", i, vector_y[i]);
}
}
static void concatenate(char p[], char q[]) {
int c;
int d;
c = 0;
while (p[c] != '\0') {
c++;
}
d = 0;
while (q[d] != '\0') {
p[c] = q[d];
d++;
c++;
}
p[c] = '\0';
}
static void compute_mean_x_y()
{
size_t i;
double tmp_x = 0.0;
double tmp_y = 0.0;
for (i = 0; i < size; i++) {
tmp_x += vector_x[i];
tmp_y += vector_y[i];
}
mean_x = tmp_x / size;
mean_y = tmp_y / size;
printf("mean_x = %lf\n", mean_x);
printf("mean_y = %lf\n", mean_y);
}
static void compute_teta_1_grad()
{
double numerator = 0.0;
double denominator = 0.0;
double tmp1 = 0.0;
double tmp2 = 0.0;
size_t i;
for (i = 0; i < size; i++) {
numerator += (vector_x[i] - mean_x) * (vector_y[i] - mean_y);
}
for (i = 0; i < size; i++) {
tmp1 = vector_x[i] - mean_x;
tmp2 = tmp1 * tmp1;
denominator += tmp2;
}
teta_1_grad = numerator / denominator;
printf("teta_1_grad = %lf\n", teta_1_grad);
}
static void compute_teta_0_intercept()
{
teta_0_intercept = mean_y - (teta_1_grad * mean_x);
printf("teta_0_intercept = %lf\n", teta_0_intercept);
}
static void compute_prediction()
{
size_t i;
for (i = 0; i < size; i++) {
vector_predict[i] = teta_0_intercept + (teta_1_grad * vector_x[i]);
printf("y^[%ld] = %lf\n", i, vector_predict[i]);
}
printf("\n");
}
static void compute_RMSE()
{
compute_prediction();
double error = 0;
size_t i;
for (i = 0; i < size; i++) {
error = (vector_predict[i] - vector_y[i]) * (vector_predict[i] - vector_y[i]);
printf("error y^[%ld] = %lf\n", i, error);
RMSE += error;
}
/* mean */
RMSE = RMSE / size;
/* square root mean */
RMSE = sqrt(RMSE);
printf("\nRMSE = %lf\n", RMSE);
}
static void compute_loss_function()
{
int input = 0;
printf("Which loss function do you want to use?\n");
printf(" 1 - RMSE\n");
scanf("%d", &input);
switch(input) {
case 1:
concatenate(cmd_gnu_0, lossname_RMSE);
compute_RMSE();
printf("\n");
break;
default:
printf("Wrong input try again\n");
compute_loss_function(size);
}
}
static void compute_r_square(size_t size)
{
double num_err = 0.0;
double den_err = 0.0;
size_t i;
for (i = 0; i < size; i++) {
num_err += (vector_y[i] - vector_predict[i]) * (vector_y[i] - vector_predict[i]);
den_err += (vector_y[i] - mean_y) * (vector_y[i] - mean_y);
}
r_square = 1 - (num_err/den_err);
printf("R_square = %lf\n", r_square);
}
static void compute_predict_for_x()
{
double x = 0.0;
printf("Please enter x value\n");
scanf("%lf", &x);
prediction = teta_0_intercept + (teta_1_grad * x);
printf("y^ if x = %lf -> %lf\n",x, prediction);
}
static void compute_max_x_y()
{
size_t i;
double tmp1= 0.0;
double tmp2= 0.0;
for (i = 0; i < size; i++) {
if (vector_x[i] > tmp1) {
tmp1 = vector_x[i];
max_x = vector_x[i];
}
if (vector_y[i] > tmp2) {
tmp2 = vector_y[i];
max_y = vector_y[i];
}
}
printf("vector_x max value %lf\n", max_x);
printf("vector_y max value %lf\n", max_y);
}
static void display_model_line()
{
sprintf(intercept, "%0.7lf", teta_0_intercept);
sprintf(grad, "%0.7lf", teta_1_grad);
sprintf(xrange, "%0.7lf", max_x + 1);
sprintf(yrange, "%0.7lf", max_y + 1);
concatenate(cmd_gnu_1, intercept);
concatenate(cmd_gnu_2, grad);
concatenate(cmd_gnu_3, xrange);
concatenate(cmd_gnu_3, "]");
concatenate(cmd_gnu_4, yrange);
concatenate(cmd_gnu_4, "]");
printf("grad = %s\n", grad);
printf("intercept = %s\n", intercept);
printf("xrange = %s\n", xrange);
printf("yrange = %s\n", yrange);
printf("cmd_gnu_0: %s\n", cmd_gnu_0);
printf("cmd_gnu_1: %s\n", cmd_gnu_1);
printf("cmd_gnu_2: %s\n", cmd_gnu_2);
printf("cmd_gnu_3: %s\n", cmd_gnu_3);
printf("cmd_gnu_4: %s\n", cmd_gnu_4);
printf("cmd_gnu_5: %s\n", cmd_gnu_5);
printf("cmd_gnu_6: %s\n", cmd_gnu_6);
/* print plot */
FILE *gnuplot_pipe = (FILE*)popen("gnuplot -persistent", "w");
FILE *temp = (FILE*)fopen("data.temp", "w");
/* create data.temp */
size_t i;
for (i = 0; i < size; i++)
{
fprintf(temp, "%f %f \n", vector_x[i], vector_y[i]);
}
/* display gnuplot */
for (i = 0; i < 7; i++)
{
fprintf(gnuplot_pipe, "%s \n", commands_gnuplot[i]);
}
}
int main(void)
{
printf("===========================================\n");
printf("INPUT DATA\n");
printf("===========================================\n");
user_input();
display_vector();
printf("\n");
printf("===========================================\n");
printf("COMPUTE MEAN X:Y, TETA_1 TETA_0\n");
printf("===========================================\n");
compute_mean_x_y();
compute_max_x_y();
compute_teta_1_grad();
compute_teta_0_intercept();
printf("\n");
printf("===========================================\n");
printf("COMPUTE LOSS FUNCTION\n");
printf("===========================================\n");
compute_loss_function();
printf("===========================================\n");
printf("COMPUTE R_square\n");
printf("===========================================\n");
compute_r_square(size);
printf("\n");
printf("===========================================\n");
printf("COMPUTE y^ according to x\n");
printf("===========================================\n");
compute_predict_for_x();
printf("\n");
printf("===========================================\n");
printf("DISPLAY LINEAR REGRESSION\n");
printf("===========================================\n");
display_model_line();
printf("\n");
return 0;
}
Look at Section 1 of this paper. This section expresses a 2D linear regression as a matrix multiplication exercise. As long as your data is well-behaved, this technique should permit you to develop a quick least squares fit.
Depending on the size of your data, it might be worthwhile to algebraically reduce the matrix multiplication to simple set of equations, thereby avoiding the need to write a matmult() function. (Be forewarned, this is completely impractical for more than 4 or 5 data points!)
The fastest, most efficient way to solve least squares, as far as I am aware, is to subtract (the gradient)/(the 2nd order gradient) from your parameter vector. (2nd order gradient = i.e. the diagonal of the Hessian.)
Here is the intuition:
Let's say you want to optimize least squares over a single parameter. This is equivalent to finding the vertex of a parabola. Then, for any random initial parameter, x0, the vertex of the loss function is located at x0 - f(1) / f(2). That's because adding - f(1) / f(2) to x will always zero out the derivative, f(1).
Side note: Implementing this in Tensorflow, the solution appeared at w0 - f(1) / f(2) / (number of weights), but I'm not sure if that's due to Tensorflow or if it's due to something else..