5f3cba02
David Mayerich
initial public co...
|
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
|
//OpenCV
#include <opencv2/opencv.hpp>
#include <stim/math/matrix.h>
#include <stim/math/constants.h>
#include <sstream>
#include "progress_thread.h"
#include <limits>
#include <chrono>
//LAPACKE support for Visual Studio
#include <complex>
#ifndef LAPACK_COMPLEX_CUSTOM
#define LAPACK_COMPLEX_CUSTOM
#define lapack_complex_float std::complex<float>
#define lapack_complex_double std::complex<double>
#endif
#include "lapacke.h"
class stim_EM : public cv::EM{
public:
cv::vector<cv::Mat> getCovs() {
return covs;
}
stim_EM(int nclusters = EM::DEFAULT_NCLUSTERS, int covMatType = EM::COV_MAT_DIAGONAL,
const cv::TermCriteria& termCrit = cv::TermCriteria(cv::TermCriteria::COUNT + cv::TermCriteria::EPS,
EM::DEFAULT_MAX_ITERS, FLT_EPSILON)) : cv::EM(nclusters, covMatType, termCrit) {
}
};
//define a structure for a multi-class GMM
class GMM{
public:
size_t K; //number of Gaussians per class
size_t F; //number of features
double t_gauss;
std::vector< double > w; //array of K weights for each Gaussian
std::vector< stim::matrix< double > > mu; //a vector storing K mean vectors of length F
//std::vector< std::vector< gmm_mat > > sigma; //(C x K) array of covariance matrices (F x F)
std::vector< stim::matrix<double> > sigma; //array of K (F x F) covariance matrices for each Gaussian
std::vector< stim::matrix<double> > sigma_i; //stores the inverse covariance matrices
std::vector< double > sqrt_tau_sigma_det; //stores sqrt(2*pi*|sigma|)
void init(){
w.resize(K); //allocate space for weights
mu.resize(K); //allocate space for means
sigma.resize(K); //allocate space for each covariance matrix
for (size_t k = 0; k < K; k++) {
mu[k] = stim::matrix<double>(F, 1);
sigma[k] = stim::matrix<double>(F, F);
}
t_gauss = 0;
}
//calculate the inverse sigma matrices
void invert_sigmas() {
sigma_i.resize(K); //allocate space for K inverse matrices
int *IPIV = (int*)malloc(sizeof(int) * F); //allocate space for the row indices
for (size_t k = 0; k < K; k++) { //for each sigma matrix
sigma_i[k] = sigma[k]; //copy the covariance matrix
LAPACKE_dgetrf(LAPACK_COL_MAJOR, (int)F, (int)F, sigma_i[k].data(), (int)F, IPIV); //perform LU factorization
LAPACKE_dgetri(LAPACK_COL_MAJOR, (int)F, sigma_i[k].data(), (int)F, IPIV); //calculate matrix inverse
}
free(IPIV);
}
void calc_sqrt_tau_sigma_det() {
sqrt_tau_sigma_det.resize(K);
for (size_t k = 0; k < K; k++) {
sqrt_tau_sigma_det[k] = sqrt(sigma[k].det() * stim::TAU);
}
}
//initialize predictors for improving calculation of responses
void init_predictors() {
invert_sigmas();
calc_sqrt_tau_sigma_det();
}
//calculate the value of a multi-variate gaussian distribution given a vector of means and a covariance matrix
double mvgauss(stim::matrix<double> x, size_t k) {
std::chrono::high_resolution_clock::time_point t0 = std::chrono::high_resolution_clock::now();
stim::matrix<double> xmu = x - mu[k];
stim::matrix<double> xmu_t = xmu.transpose();
stim::matrix<double> xmu_t_sigma_i = xmu_t * sigma_i[k];
stim::matrix<double> xmu_t_sigma_i_xmu = xmu_t_sigma_i * xmu;
double a = -0.5 * xmu_t_sigma_i_xmu(0, 0);
double numer = exp(a);
stim::matrix<double> tau_sigma = sigma[k] * stim::TAU;
double determinant = tau_sigma.det();
double denom = sqrt(determinant);
std::chrono::high_resolution_clock::time_point t1 = std::chrono::high_resolution_clock::now();
t_gauss += std::chrono::duration_cast< std::chrono::duration<double> >(t1 - t0).count();
return numer / denom;
}
double mvgauss(double* x, size_t k, double* scratch) {
std::chrono::high_resolution_clock::time_point t0 = std::chrono::high_resolution_clock::now();
for (size_t f = 0; f < F; f++)
scratch[f] = x[f] - mu[k](f, 0);
stim::matrix<double> xmu(F, 1, scratch);
stim::matrix<double> xmu_t(1, F, scratch);
stim::matrix<double> xmu_t_sigma_i = xmu_t * sigma_i[k];
stim::matrix<double> xmu_t_sigma_i_xmu = xmu_t_sigma_i * xmu;
double a = -0.5 * xmu_t_sigma_i_xmu(0, 0);
double numer = exp(a);
std::chrono::high_resolution_clock::time_point t1 = std::chrono::high_resolution_clock::now();
t_gauss += std::chrono::duration_cast< std::chrono::duration<double> >(t1 - t0).count();
return numer / sqrt_tau_sigma_det[k];
}
/// returns the probability density of the membership of v in all K clusters
std::vector<double> G(stim::matrix<double> x) {
std::vector<double> result(K); //allocate space for all K probabilities
for (size_t k = 0; k < K; k++) { //for each gaussian
result[k] = mvgauss(x, k);
}
return result;
}
/// Calculate the response to x among all K clusters given pointers to pre-allocated arrays
void G(double* x, double* r) {
double* scratch = (double*)malloc(F * sizeof(double));
for (size_t k = 0; k < K; k++) { //for each gaussian
r[k] = mvgauss(x, k, scratch);
}
free(scratch);
}
/// Return the cluster most closely corresponding to the input vector x
size_t get_cluster(stim::matrix<double> x) {
size_t cluster; //stores the cluster ID
std::vector<double> posteriors = G(x);
double largest = posteriors[0];
for (size_t k = 0; k < K; k++) {
if (posteriors[k] >= largest) {
largest = posteriors[k];
cluster = k;
}
}
return cluster;
}
///Return the posterior probability of the vector x based on the current Gaussian mixture model
double P(stim::matrix<double> x) {
std::vector<double> posteriors = G(x);
double p = 0;
for (size_t k = 0; k < K; k++) {
p += w[k] * posteriors[k]; //calculate the weighted sum of all Gaussian functions
}
return p;
}
double P(double* x) {
double* posteriors = (double*)malloc(K * sizeof(double));
G(x, posteriors);
double p = 0;
for (size_t k = 0; k < K; k++) {
p += w[k] * posteriors[k]; //calculate the weighted sum of all Gaussian functions
}
return p;
}
public:
GMM() {
K = 0;
F = 0;
}
GMM(size_t clusters, size_t features){
K = clusters;
F = features;
init();
}
void set(const cv::Mat weights, const cv::Mat means, const std::vector<cv::Mat> cov) {
for (size_t k = 0; k < K; k++)
w[k] = weights.at<double>(0, (int)k);
for (size_t k = 0; k < K; k++)
for (size_t f = 0; f < F; f++)
mu[k](f, 0) = means.at<double>((int)k, (int)f);
for (size_t k = 0; k < K; k++) {
for (size_t fi = 0; fi < F; fi++) {
for (size_t fj = 0; fj < F; fj++) {
sigma[k](fi, fj) = cov[k].at<double>((int)fi, (int)fj);
}
}
}
init_predictors(); //calculate the inverse covariance matrices
}
std::string str() {
std::stringstream ss;
ss << "weights:" << std::endl;
for (size_t k = 0; k < K; k++)
ss << " " << w[k] << std::endl;
ss << std::endl << "centers:" << std::endl;
for (size_t k = 0; k < K; k++)
ss << mu[k].toStr() << std::endl;
ss << std::endl << "covariances:" << std::endl;
for (size_t k = 0; k < K; k++)
ss << sigma[k].toStr() << std::endl;
return ss.str();
}
void save(std::ostream& out) {
out << K << std::endl; //save the number of clusters
out << F << std::endl; //save the number of features
for (size_t k = 0; k < K; k++)
out << std::fixed << w[k] << std::endl;
for (size_t k = 0; k < K; k++)
out << mu[k].csv() << std::endl;
for (size_t k = 0; k < K; k++)
out << sigma[k].csv() << std::endl;
}
void save(std::string filename) {
std::ofstream outfile(filename);
int digits = std::numeric_limits<double>::max_digits10;
outfile.precision(digits);
save(outfile);
outfile.close();
}
//load a GMM
void load(std::istream& in) {
in >> K; //load the number of clusters
in >> F; //load the number of features
init();
for (size_t k = 0; k < K; k++)
in >> w[k];
for (size_t k = 0; k < K; k++)
mu[k].csv(in);
for (size_t k = 0; k < K; k++)
sigma[k].csv(in);
init_predictors(); //calculate the inverse covariance matrices
}
void load(std::string filename) {
std::ifstream infile(filename);
load(infile);
infile.close();
}
};
/// Multi-class supervised GMM
class multiGMM {
public:
size_t C; //number of classes
std::vector<GMM> gmms; //vector of Gaussian Mixture models
/// Generate an empty GMM for each class
void init() {
for (size_t c = 0; c < C; c++) {
gmms.resize(C);
}
}
multiGMM(size_t classes) {
C = classes; //store the number of classes
init();
}
//get the class that most likely corresponds to x
size_t get_class(stim::matrix<double> x) {
double p0;
size_t c_p = 0; //stores the most likely class label
double p = gmms[0].P(x); //get the posterior probability of class 0
for (size_t c = 1; c < C; c++) { //for each class
p0 = gmms[c].P(x); //get the posterior probability of membership given x
if (p0 > p) { //if the new class is most likely
p = p0; //update the maximum probability
c_p = c; //update the class ID
}
}
return c_p;
}
void save(std::string filename) {
std::ofstream outfile(filename); //open an output file stream
if (outfile) {
int digits = std::numeric_limits<double>::max_digits10;
outfile.precision(digits);
outfile << C << std::endl; //save the number of classes
for (size_t c = 0; c < C; c++) {
gmms[c].save(outfile); //save each individual GMM
}
outfile.close();
}
else {
std::cout << "ERROR creating GMM file " << filename << std::endl;
exit(1);
}
}
bool load(std::string filename) {
std::ifstream infile(filename); //open the input file
if (!infile) return false;
infile >> C; //load the number of classes
gmms.resize(C); //resize the GMM array to match the number of classes
for (size_t c = 0; c < C; c++) //load each GMM (one per class)
gmms[c].load(infile);
return true;
}
};
/// trains a single Gaussian Mixture model using expectation maximization in OpenCV
GMM train_gmm(cv::Mat &F, int k, int attempts, int iters, double epsilon){
GMM new_gmm(k, F.cols); //create a new GMM classifier
stim_EM em(k, cv::EM::COV_MAT_DIAGONAL, cv::TermCriteria( CV_TERMCRIT_EPS+CV_TERMCRIT_ITER, iters, epsilon));
if(!em.train(F)) {
std::cout << "ERROR training GMM" << std::endl;
exit(1);
}
size_t nc = em.get<int>("nclusters");
cv::Mat output;
cv::Mat means = em.get<cv::Mat>("means");
cv::Mat weights = em.get<cv::Mat>("weights");
cv::vector<cv::Mat> covs = em.getCovs();
new_gmm.set(weights, means, covs);
return new_gmm;
}
//Predict a set of classes based on given centroid vectors
std::vector< stim::image<unsigned char> > predict_gmm(stim::envi* E, multiGMM* gmm, std::vector< stim::image<float> >& responses, unsigned char* MASK = NULL){
size_t nC = gmm->C; //get the number of classes
if (nC == 1)
nC = gmm->gmms[0].K; //if there is only one GMM, classify based on clusters
size_t X = E->header.samples; //store ENVI file size parameters
size_t Y = E->header.lines;
size_t B = E->header.bands;
size_t XY = E->header.samples * E->header.lines;
size_t tP = 0; //calculate the total number of pixels
if(MASK){
for(size_t xy = 0; xy < XY; xy++){
if(MASK[xy]) tP++;
}
}
else
tP = X * Y;
std::vector< stim::image<unsigned char> > C; //create an array of mask images
C.resize(nC);
responses.resize(nC); //allocate space for the response images
for(size_t c = 0; c < nC; c++){ //for each class mask
C[c] = stim::image<unsigned char>(X, Y, 1); //allocate space for the mask
memset(C[c].data(), 0, X * Y * sizeof(unsigned char)); //initialize all of the pixels to zero
responses[c] = stim::image<float>(X, Y, 1); //allocate space for the response image
memset(responses[c].data(), 0, X * Y * sizeof(float)); //initialize the response image to zero
}
double progress = 0; //initialize the progress bar variable
std::thread t1(progressbar_thread, &progress); //start the progress bar thread
size_t t = 0;
double* spectrum = (double*)malloc(sizeof(double) * B); //allocate space to hold a spectrum
double gm, maxgm;
size_t maxc;
for(size_t p = 0; p < XY; p++){ //for each pixel
if(!MASK || MASK[p] > 0){
E->spectrum<double>(spectrum, p); //get the spectrum at pixel p
maxc = 0;
for (size_t c = 0; c < nC; c++) {
gm = gmm->gmms[c].P(spectrum); //evaluate the posterior for class c
responses[c].data()[p] = (float)gm;
if (c == 0) maxgm = gm;
else if (gm > maxgm) {
maxgm = gm;
maxc = c;
}
}
C[maxc].data()[p] = 255;
t++;
progress = (double)(t+1) / (double)(tP) * 100.0; //update the progress bar variable
}
}
t1.join(); //finish the progress bar thread
for (size_t c = 0; c < gmm->gmms.size(); c++) {
std::cout << "gauss-time (" << c << "): " << gmm->gmms[c].t_gauss << std::endl;
}
return C;
}
|