| 1 |
// |
| 2 |
// Created by Kai Zhao on 12/9/19. |
| 3 |
// |
| 4 |
|
| 5 |
#ifndef _SZ_KMEANS_UTIL |
| 6 |
#define _SZ_KMEANS_UTIL |
| 7 |
|
| 8 |
#include <algorithm> |
| 9 |
#include <functional> |
| 10 |
#include <numeric> |
| 11 |
#include <unordered_map> |
| 12 |
#include <random> |
| 13 |
#include <vector> |
| 14 |
#include <unordered_set> |
| 15 |
#include <SZ3/utils/Timer.hpp> |
| 16 |
#include <SZ3/def.hpp> |
| 17 |
|
| 18 |
|
| 19 |
namespace SZ { |
| 20 |
|
| 21 |
/* |
| 22 |
* Internal implementation of the SMAWK algorithm. |
| 23 |
*/ |
| 24 |
template<typename T> |
| 25 |
void _smawk( |
| 26 |
const std::vector<size_t> &rows, |
| 27 |
const std::vector<size_t> &cols, |
| 28 |
const std::function<T(size_t, size_t)> &lookup, |
| 29 |
std::vector<size_t> *result) { |
| 30 |
// Recursion base case |
| 31 |
if (rows.size() == 0) return; |
| 32 |
|
| 33 |
// ******************************** |
| 34 |
// * REDUCE |
| 35 |
// ******************************** |
| 36 |
|
| 37 |
std::vector<size_t> _cols; // Stack of surviving columns |
| 38 |
for (size_t col : cols) { |
| 39 |
while (true) { |
| 40 |
if (_cols.size() == 0) break; |
| 41 |
size_t row = rows[_cols.size() - 1]; |
| 42 |
if (lookup(row, col) >= lookup(row, _cols.back())) |
| 43 |
break; |
| 44 |
_cols.pop_back(); |
| 45 |
} |
| 46 |
if (_cols.size() < rows.size()) |
| 47 |
_cols.push_back(col); |
| 48 |
} |
| 49 |
|
| 50 |
// Call recursively on odd-indexed rows |
| 51 |
std::vector<size_t> odd_rows; |
| 52 |
for (size_t i = 1; i < rows.size(); i += 2) { |
| 53 |
odd_rows.push_back(rows[i]); |
| 54 |
} |
| 55 |
_smawk(odd_rows, _cols, lookup, result); |
| 56 |
|
| 57 |
std::unordered_map<size_t, size_t> col_idx_lookup; |
| 58 |
for (size_t idx = 0; idx < _cols.size(); ++idx) { |
| 59 |
col_idx_lookup[_cols[idx]] = idx; |
| 60 |
} |
| 61 |
|
| 62 |
// ******************************** |
| 63 |
// * INTERPOLATE |
| 64 |
// ******************************** |
| 65 |
|
| 66 |
// Fill-in even-indexed rows |
| 67 |
size_t start = 0; |
| 68 |
for (size_t r = 0; r < rows.size(); r += 2) { |
| 69 |
size_t row = rows[r]; |
| 70 |
size_t stop = _cols.size() - 1; |
| 71 |
if (r < rows.size() - 1) |
| 72 |
stop = col_idx_lookup[(*result)[rows[r + 1]]]; |
| 73 |
size_t argmin = _cols[start]; |
| 74 |
T min = lookup(row, argmin); |
| 75 |
for (size_t c = start + 1; c <= stop; ++c) { |
| 76 |
T value = lookup(row, _cols[c]); |
| 77 |
if (c == start || value < min) { |
| 78 |
argmin = _cols[c]; |
| 79 |
min = value; |
| 80 |
} |
| 81 |
} |
| 82 |
(*result)[row] = argmin; |
| 83 |
start = stop; |
| 84 |
} |
| 85 |
} |
| 86 |
|
| 87 |
/* |
| 88 |
* Interface for the SMAWK algorithm, for finding the minimum value in each row |
| 89 |
* of an implicitly-defined totally monotone matrix. |
| 90 |
*/ |
| 91 |
template<typename T> |
| 92 |
std::vector<size_t> smawk( |
| 93 |
const size_t num_rows, |
| 94 |
const size_t num_cols, |
| 95 |
const std::function<T(size_t, size_t)> &lookup) { |
| 96 |
std::vector<size_t> result; |
| 97 |
result.resize(num_rows); |
| 98 |
std::vector<size_t> rows(num_rows); |
| 99 |
iota(begin(rows), end(rows), 0); |
| 100 |
std::vector<size_t> cols(num_cols); |
| 101 |
iota(begin(cols), end(cols), 0); |
| 102 |
_smawk<T>(rows, cols, lookup, &result); |
| 103 |
return result; |
| 104 |
} |
| 105 |
|
| 106 |
/* |
| 107 |
* Calculates cluster costs in O(1) using prefix sum arrays. |
| 108 |
*/ |
| 109 |
template<class DT> |
| 110 |
class CostCalculator { |
| 111 |
std::vector<double> cumsum; |
| 112 |
std::vector<double> cumsum2; |
| 113 |
|
| 114 |
public: |
| 115 |
CostCalculator(const std::vector<DT> &vec, size_t n) { |
| 116 |
cumsum.push_back(0.0); |
| 117 |
cumsum2.push_back(0.0); |
| 118 |
for (size_t i = 0; i < n; ++i) { |
| 119 |
double x = vec[i]; |
| 120 |
cumsum.push_back(x + cumsum[i]); |
| 121 |
cumsum2.push_back(x * x + cumsum2[i]); |
| 122 |
} |
| 123 |
} |
| 124 |
|
| 125 |
double calc(size_t i, size_t j) { |
| 126 |
if (j < i) return 0.0; |
| 127 |
double mu = (cumsum[j + 1] - cumsum[i]) / (j - i + 1); |
| 128 |
double result = cumsum2[j + 1] - cumsum2[i]; |
| 129 |
result += (j - i + 1) * (mu * mu); |
| 130 |
result -= (2 * mu) * (cumsum[j + 1] - cumsum[i]); |
| 131 |
return result; |
| 132 |
} |
| 133 |
}; |
| 134 |
|
| 135 |
template<typename T> |
| 136 |
class Matrix { |
| 137 |
std::vector<T> data; |
| 138 |
size_t num_rows; |
| 139 |
size_t num_cols; |
| 140 |
|
| 141 |
public: |
| 142 |
Matrix(size_t num_rows, size_t num_cols) { |
| 143 |
this->num_rows = num_rows; |
| 144 |
this->num_cols = num_cols; |
| 145 |
data.resize(num_rows * num_cols); |
| 146 |
} |
| 147 |
|
| 148 |
inline T get(size_t i, size_t j) { |
| 149 |
return data[i * num_cols + j]; |
| 150 |
} |
| 151 |
|
| 152 |
inline void set(size_t i, size_t j, T value) { |
| 153 |
data[i * num_cols + j] = value; |
| 154 |
} |
| 155 |
}; |
| 156 |
|
| 157 |
template<class DT> |
| 158 |
void cluster( |
| 159 |
DT *array, |
| 160 |
size_t n, |
| 161 |
int &k, |
| 162 |
size_t *clusters, |
| 163 |
DT *centroids) { |
| 164 |
// *************************************************** |
| 165 |
// * Sort input array and save info for de-sorting |
| 166 |
// *************************************************** |
| 167 |
|
| 168 |
std::vector<size_t> sort_idxs(n); |
| 169 |
iota(sort_idxs.begin(), sort_idxs.end(), 0); |
| 170 |
sort( |
| 171 |
sort_idxs.begin(), |
| 172 |
sort_idxs.end(), |
| 173 |
[&array](size_t a, size_t b) { return array[a] < array[b]; }); |
| 174 |
// vector<size_t> undo_sort_lookup(n); |
| 175 |
std::vector<DT> sorted_array(n); |
| 176 |
for (size_t i = 0; i < n; ++i) { |
| 177 |
sorted_array[i] = array[sort_idxs[i]]; |
| 178 |
// undo_sort_lookup[sort_idxs[i]] = i; |
| 179 |
} |
| 180 |
|
| 181 |
// *************************************************** |
| 182 |
// * Set D and T using dynamic programming algorithm |
| 183 |
// *************************************************** |
| 184 |
|
| 185 |
// Algorithm as presented in section 2.2 of (Gronlund et al., 2017). |
| 186 |
|
| 187 |
CostCalculator<DT> cost_calculator(sorted_array, n); |
| 188 |
Matrix<DT> D(k, n); |
| 189 |
Matrix<size_t> T(k, n); |
| 190 |
|
| 191 |
for (size_t i = 0; i < n; ++i) { |
| 192 |
D.set(0, i, cost_calculator.calc(0, i)); |
| 193 |
T.set(0, i, 0); |
| 194 |
} |
| 195 |
|
| 196 |
double ratio_avg = 0; |
| 197 |
bool findk = false; |
| 198 |
size_t bestk = 0; |
| 199 |
for (size_t k_ = 1; k_ < k; ++k_) { |
| 200 |
auto C = [&D, &k_, &cost_calculator](size_t i, size_t j) -> DT { |
| 201 |
size_t col = i < j - 1 ? i : j - 1; |
| 202 |
return D.get(k_ - 1, col) + cost_calculator.calc(j, i); |
| 203 |
}; |
| 204 |
std::vector<size_t> row_argmins = smawk<DT>(n, n, C); |
| 205 |
for (size_t i = 0; i < row_argmins.size(); ++i) { |
| 206 |
size_t argmin = row_argmins[i]; |
| 207 |
DT min = C(i, argmin); |
| 208 |
D.set(k_, i, min); |
| 209 |
T.set(k_, i, argmin); |
| 210 |
} |
| 211 |
float ratio = D.get(k_ - 1, n - 1) / D.get(k_, n - 1); |
| 212 |
ratio_avg = (ratio_avg * (k_ - 1) + ratio) / (k_); |
| 213 |
// std::cout << k_ + 1 << " , " << D.get(k_, n - 1) << " , " << ratio << " , " << ratio_avg << std::endl; |
| 214 |
if (ratio / ratio_avg > 1.5) { |
| 215 |
bestk = k_ + 1; |
| 216 |
findk = true; |
| 217 |
} else { |
| 218 |
if (findk) { |
| 219 |
break; |
| 220 |
} |
| 221 |
} |
| 222 |
} |
| 223 |
|
| 224 |
if (!findk) { |
| 225 |
return; |
| 226 |
} |
| 227 |
k = bestk; |
| 228 |
std::cout << "# groups = " << k << std::endl; |
| 229 |
|
| 230 |
// *************************************************** |
| 231 |
// * Extract cluster assignments by backtracking |
| 232 |
// *************************************************** |
| 233 |
|
| 234 |
// Note: This step requires O(kn) memory usage due to saving the entire |
| 235 |
// T matrix. However, it can be modified so that the memory usage is O(n). |
| 236 |
// D and T would not need to be retained in full (D already doesn't need |
| 237 |
// to be fully retained, although it currently is). |
| 238 |
// Details are in section 3 of (Grønlund et al., 2017). |
| 239 |
|
| 240 |
// vector<DT> sorted_clusters(n); |
| 241 |
|
| 242 |
size_t t = n; |
| 243 |
size_t k_ = k - 1; |
| 244 |
size_t n_ = n - 1; |
| 245 |
// The do/while loop was used in place of: |
| 246 |
// for (k_ = k - 1; k_ >= 0; --k_) |
| 247 |
// to avoid wraparound of an unsigned type. |
| 248 |
do { |
| 249 |
size_t t_ = t; |
| 250 |
t = T.get(k_, n_); |
| 251 |
DT centroid = 0.0; |
| 252 |
for (size_t i = t; i < t_; ++i) { |
| 253 |
// sorted_clusters[i] = k_; |
| 254 |
centroid += (sorted_array[i] - centroid) / (i - t + 1); |
| 255 |
} |
| 256 |
centroids[k_] = centroid; |
| 257 |
k_ -= 1; |
| 258 |
n_ = t - 1; |
| 259 |
} while (t > 0); |
| 260 |
|
| 261 |
// *************************************************** |
| 262 |
// * Order cluster assignments to match de-sorted |
| 263 |
// * ordering |
| 264 |
// *************************************************** |
| 265 |
|
| 266 |
// for (size_t i = 0; i < n; ++i) { |
| 267 |
// clusters[i] = sorted_clusters[undo_sort_lookup[i]]; |
| 268 |
// } |
| 269 |
} |
| 270 |
|
| 271 |
template<class T> |
| 272 |
int f(T data, double start_position, double offset) { |
| 273 |
return round((data - start_position) / offset); |
| 274 |
} |
| 275 |
|
| 276 |
template<class T> |
| 277 |
int f1(T data, T *boundary, int n, double start_position, double offset) { |
| 278 |
return round((data - start_position) / offset); |
| 279 |
} |
| 280 |
|
| 281 |
template<class T> |
| 282 |
int f2(T data, T *boundary, int n, double start_position, double offset) { |
| 283 |
int low = 0, high = n; // numElems is the size of the array i.e arr.size() |
| 284 |
while (low != high) { |
| 285 |
int mid = (low + high) / 2; // Or a fancy way to avoid int overflow |
| 286 |
if (boundary[mid] <= data) { |
| 287 |
low = mid + 1; |
| 288 |
} else { |
| 289 |
high = mid; |
| 290 |
} |
| 291 |
} |
| 292 |
return high; |
| 293 |
} |
| 294 |
|
| 295 |
template<class T> |
| 296 |
int f3(T data, T *boundary, int n, double start_position, double offset) { |
| 297 |
for (int i = 1; i < n; i++) { |
| 298 |
if (boundary[i] > data) { |
| 299 |
return i - 1; |
| 300 |
} |
| 301 |
} |
| 302 |
return n - 1; |
| 303 |
} |
| 304 |
|
| 305 |
|
| 306 |
template<class T> |
| 307 |
void get_cluster(T *data, size_t num, float &level_start, float &level_offset, int &level_num, |
| 308 |
size_t sample_num) { |
| 309 |
T max = *std::max_element(data, data + num); |
| 310 |
T min = *std::min_element(data, data + num); |
| 311 |
SZ::Timer timer; |
| 312 |
timer.start(); |
| 313 |
std::vector<T> sample; |
| 314 |
if (num == sample_num) { |
| 315 |
sample = std::vector<T>(data, data + num); |
| 316 |
} else { |
| 317 |
sample.reserve(sample_num); |
| 318 |
std::random_device rd; //Will be used to obtain a seed for the random number engine |
| 319 |
std::mt19937 gen(rd()); //Standard mersenne_twister_engine seeded with rd() |
| 320 |
// std::uniform_int_distribution<> dis(0, 2 * sample_rate); |
| 321 |
size_t input_idx = 0; |
| 322 |
// for (int i = 0; i < num / sample_rate; i++) { |
| 323 |
// input_idx = (input_idx + dis(gen)) % num; |
| 324 |
// std::cout << input_idx << " "; |
| 325 |
// sample[i] = input[input_idx]; |
| 326 |
// } |
| 327 |
// std::cout << std::endl; |
| 328 |
std::uniform_int_distribution<> dis2(0, num); |
| 329 |
std::unordered_set<size_t> sampledkeys; |
| 330 |
printf("total_num=%lu, sample_num=%lu\n", num, sample_num); |
| 331 |
for (size_t i = 0; i < sample_num; i++) { |
| 332 |
do { |
| 333 |
input_idx = dis2(gen); |
| 334 |
} while (sampledkeys.find(input_idx) != sampledkeys.end()); |
| 335 |
// std::cout << input_idx << " "; |
| 336 |
sample[i] = data[input_idx]; |
| 337 |
} |
| 338 |
// std::cout << std::endl; |
| 339 |
|
| 340 |
// std::sample(input.begin(), input.end(), |
| 341 |
// sample.begin(), |
| 342 |
// num / sample_rate, |
| 343 |
// std::mt19937{std::random_device{}()}); |
| 344 |
} |
| 345 |
|
| 346 |
timer.stop("random sample"); |
| 347 |
// sample = input; |
| 348 |
std::vector<size_t> idx(num); |
| 349 |
|
| 350 |
timer.start(); |
| 351 |
int k = 150; |
| 352 |
std::vector<float> cents(k); |
| 353 |
cluster(sample.data(), sample_num, k, idx.data(), cents.data()); |
| 354 |
// cluster(input.get(), num, 16, idx.data(), cents.data()); |
| 355 |
timer.stop("kmeans1d"); |
| 356 |
if (k == 150) { |
| 357 |
std::cout << "No clusters are found." << std::endl; |
| 358 |
level_num = 0; |
| 359 |
return; |
| 360 |
// exit(0); |
| 361 |
} |
| 362 |
|
| 363 |
// std::cout << "centers : "; |
| 364 |
// for (size_t i = 0; i < k; i++) { |
| 365 |
// std::cout << cents[i] << " "; |
| 366 |
// } |
| 367 |
// std::cout << std::endl; |
| 368 |
|
| 369 |
// std::cout << "center diff : "; |
| 370 |
// for (size_t i = 1; i < k; i++) { |
| 371 |
// std::cout << cents[i] - cents[i - 1] << " "; |
| 372 |
// } |
| 373 |
// std::cout << std::endl; |
| 374 |
|
| 375 |
// std::vector<float> boundary(k); |
| 376 |
// boundary[0] = std::numeric_limits<float>::min(); |
| 377 |
// for (size_t i = 1; i < k; i++) { |
| 378 |
// boundary[i] = (cents[i - 1] + cents[i]) / 2; |
| 379 |
// } |
| 380 |
|
| 381 |
level_offset = (cents[k - 1] - cents[0]) / (k - 1); |
| 382 |
level_start = cents[0]; |
| 383 |
for (size_t i = 1; i < k; i++) { |
| 384 |
level_start += cents[i] - i * level_offset; |
| 385 |
} |
| 386 |
level_start /= k; |
| 387 |
level_num = f(max, level_start, level_offset) + 1; |
| 388 |
} |
| 389 |
} |
| 390 |
|
| 391 |
#endif |