1 | // Copyright (C) 2015 Conrad Sanderson |
---|
2 | // Copyright (C) 2015 NICTA (www.nicta.com.au) |
---|
3 | // |
---|
4 | // This Source Code Form is subject to the terms of the Mozilla Public |
---|
5 | // License, v. 2.0. If a copy of the MPL was not distributed with this |
---|
6 | // file, You can obtain one at http://mozilla.org/MPL/2.0/. |
---|
7 | |
---|
8 | |
---|
9 | //! \addtogroup fn_svds |
---|
10 | //! @{ |
---|
11 | |
---|
12 | |
---|
13 | template<typename T1> |
---|
14 | inline |
---|
15 | bool |
---|
16 | svds_helper |
---|
17 | ( |
---|
18 | Mat<typename T1::elem_type>& U, |
---|
19 | Col<typename T1::pod_type >& S, |
---|
20 | Mat<typename T1::elem_type>& V, |
---|
21 | const SpBase<typename T1::elem_type,T1>& X, |
---|
22 | const uword k, |
---|
23 | const typename T1::pod_type tol, |
---|
24 | const bool calc_UV, |
---|
25 | const typename arma_real_only<typename T1::elem_type>::result* junk = 0 |
---|
26 | ) |
---|
27 | { |
---|
28 | arma_extra_debug_sigprint(); |
---|
29 | arma_ignore(junk); |
---|
30 | |
---|
31 | typedef typename T1::elem_type eT; |
---|
32 | typedef typename T1::pod_type T; |
---|
33 | |
---|
34 | if(arma_config::arpack == false) |
---|
35 | { |
---|
36 | arma_stop("svds(): use of ARPACK needs to be enabled"); |
---|
37 | return false; |
---|
38 | } |
---|
39 | |
---|
40 | arma_debug_check |
---|
41 | ( |
---|
42 | ( ((void*)(&U) == (void*)(&S)) || (&U == &V) || ((void*)(&S) == (void*)(&V)) ), |
---|
43 | "svds(): two or more output objects are the same object" |
---|
44 | ); |
---|
45 | |
---|
46 | arma_debug_check( (tol < T(0)), "svds(): tol must be >= 0" ); |
---|
47 | |
---|
48 | const unwrap_spmat<T1> tmp(X.get_ref()); |
---|
49 | const SpMat<eT>& A = tmp.M; |
---|
50 | |
---|
51 | const uword kk = (std::min)( (std::min)(A.n_rows, A.n_cols), k ); |
---|
52 | |
---|
53 | const T A_max = (A.n_nonzero > 0) ? T(max(abs(Col<eT>(const_cast<eT*>(A.values), A.n_nonzero, false)))) : T(0); |
---|
54 | |
---|
55 | if(A_max == T(0)) |
---|
56 | { |
---|
57 | // TODO: use reset instead ? |
---|
58 | S.zeros(kk); |
---|
59 | |
---|
60 | if(calc_UV) |
---|
61 | { |
---|
62 | U.eye(A.n_rows, kk); |
---|
63 | V.eye(A.n_cols, kk); |
---|
64 | } |
---|
65 | } |
---|
66 | else |
---|
67 | { |
---|
68 | SpMat<eT> C( (A.n_rows + A.n_cols), (A.n_rows + A.n_cols) ); |
---|
69 | |
---|
70 | SpMat<eT> B = A / A_max; |
---|
71 | SpMat<eT> Bt = B.t(); |
---|
72 | |
---|
73 | C(0, A.n_rows, size(B) ) = B; |
---|
74 | C(A.n_rows, 0, size(Bt)) = Bt; |
---|
75 | |
---|
76 | Bt.reset(); |
---|
77 | B.reset(); |
---|
78 | |
---|
79 | Col<eT> eigval; |
---|
80 | Mat<eT> eigvec; |
---|
81 | |
---|
82 | const bool status = sp_auxlib::eigs_sym(eigval, eigvec, C, kk, "la", (tol / Datum<T>::sqrt2)); |
---|
83 | |
---|
84 | if(status == false) |
---|
85 | { |
---|
86 | U.reset(); |
---|
87 | S.reset(); |
---|
88 | V.reset(); |
---|
89 | |
---|
90 | return false; |
---|
91 | } |
---|
92 | |
---|
93 | const T A_norm = max(eigval); |
---|
94 | |
---|
95 | const T tol2 = tol / Datum<T>::sqrt2 * A_norm; |
---|
96 | |
---|
97 | uvec indices = find(eigval > tol2); |
---|
98 | |
---|
99 | if(indices.n_elem > kk) |
---|
100 | { |
---|
101 | indices = indices.subvec(0,kk-1); |
---|
102 | } |
---|
103 | else |
---|
104 | if(indices.n_elem < kk) |
---|
105 | { |
---|
106 | const uvec indices2 = find(abs(eigval) <= tol2); |
---|
107 | |
---|
108 | const uword N_extra = (std::min)( indices2.n_elem, (kk - indices.n_elem) ); |
---|
109 | |
---|
110 | if(N_extra > 0) { indices = join_cols(indices, indices2.subvec(0,N_extra-1)); } |
---|
111 | } |
---|
112 | |
---|
113 | const uvec sorted_indices = sort_index(eigval, "descend"); |
---|
114 | |
---|
115 | S = eigval.elem(sorted_indices); S *= A_max; |
---|
116 | |
---|
117 | if(calc_UV) |
---|
118 | { |
---|
119 | uvec U_row_indices(A.n_rows); for(uword i=0; i < A.n_rows; ++i) { U_row_indices[i] = i; } |
---|
120 | uvec V_row_indices(A.n_cols); for(uword i=0; i < A.n_cols; ++i) { V_row_indices[i] = i + A.n_rows; } |
---|
121 | |
---|
122 | U = Datum<T>::sqrt2 * eigvec(U_row_indices, sorted_indices); |
---|
123 | V = Datum<T>::sqrt2 * eigvec(V_row_indices, sorted_indices); |
---|
124 | } |
---|
125 | } |
---|
126 | |
---|
127 | arma_debug_warn( (S.n_elem < k), "svds(): warning: found fewer singular values than specified" ); |
---|
128 | |
---|
129 | return true; |
---|
130 | } |
---|
131 | |
---|
132 | |
---|
133 | |
---|
134 | template<typename T1> |
---|
135 | inline |
---|
136 | bool |
---|
137 | svds_helper |
---|
138 | ( |
---|
139 | Mat<typename T1::elem_type>& U, |
---|
140 | Col<typename T1::pod_type >& S, |
---|
141 | Mat<typename T1::elem_type>& V, |
---|
142 | const SpBase<typename T1::elem_type,T1>& X, |
---|
143 | const uword k, |
---|
144 | const typename T1::pod_type tol, |
---|
145 | const bool calc_UV, |
---|
146 | const typename arma_cx_only<typename T1::elem_type>::result* junk = 0 |
---|
147 | ) |
---|
148 | { |
---|
149 | arma_extra_debug_sigprint(); |
---|
150 | arma_ignore(junk); |
---|
151 | |
---|
152 | typedef typename T1::elem_type eT; |
---|
153 | typedef typename T1::pod_type T; |
---|
154 | |
---|
155 | if(arma_config::arpack == false) |
---|
156 | { |
---|
157 | arma_stop("svds(): use of ARPACK needs to be enabled"); |
---|
158 | return false; |
---|
159 | } |
---|
160 | |
---|
161 | arma_debug_check |
---|
162 | ( |
---|
163 | ( ((void*)(&U) == (void*)(&S)) || (&U == &V) || ((void*)(&S) == (void*)(&V)) ), |
---|
164 | "svds(): two or more output objects are the same object" |
---|
165 | ); |
---|
166 | |
---|
167 | arma_debug_check( (tol < T(0)), "svds(): tol must be >= 0" ); |
---|
168 | |
---|
169 | const unwrap_spmat<T1> tmp(X.get_ref()); |
---|
170 | const SpMat<eT>& A = tmp.M; |
---|
171 | |
---|
172 | const uword kk = (std::min)( (std::min)(A.n_rows, A.n_cols), k ); |
---|
173 | |
---|
174 | const T A_max = (A.n_nonzero > 0) ? T(max(abs(Col<eT>(const_cast<eT*>(A.values), A.n_nonzero, false)))) : T(0); |
---|
175 | |
---|
176 | if(A_max == T(0)) |
---|
177 | { |
---|
178 | // TODO: use reset instead ? |
---|
179 | S.zeros(kk); |
---|
180 | |
---|
181 | if(calc_UV) |
---|
182 | { |
---|
183 | U.eye(A.n_rows, kk); |
---|
184 | V.eye(A.n_cols, kk); |
---|
185 | } |
---|
186 | } |
---|
187 | else |
---|
188 | { |
---|
189 | SpMat<eT> C( (A.n_rows + A.n_cols), (A.n_rows + A.n_cols) ); |
---|
190 | |
---|
191 | SpMat<eT> B = A / A_max; |
---|
192 | SpMat<eT> Bt = B.t(); |
---|
193 | |
---|
194 | C(0, A.n_rows, size(B) ) = B; |
---|
195 | C(A.n_rows, 0, size(Bt)) = Bt; |
---|
196 | |
---|
197 | Bt.reset(); |
---|
198 | B.reset(); |
---|
199 | |
---|
200 | Col<eT> eigval_tmp; |
---|
201 | Mat<eT> eigvec; |
---|
202 | |
---|
203 | const bool status = sp_auxlib::eigs_gen(eigval_tmp, eigvec, C, kk, "lr", (tol / Datum<T>::sqrt2)); |
---|
204 | |
---|
205 | if(status == false) |
---|
206 | { |
---|
207 | U.reset(); |
---|
208 | S.reset(); |
---|
209 | V.reset(); |
---|
210 | arma_bad("svds(): failed to converge", false); |
---|
211 | |
---|
212 | return false; |
---|
213 | } |
---|
214 | |
---|
215 | const Col<T> eigval = real(eigval_tmp); |
---|
216 | |
---|
217 | const T A_norm = max(eigval); |
---|
218 | |
---|
219 | const T tol2 = tol / Datum<T>::sqrt2 * A_norm; |
---|
220 | |
---|
221 | uvec indices = find(eigval > tol2); |
---|
222 | |
---|
223 | if(indices.n_elem > kk) |
---|
224 | { |
---|
225 | indices = indices.subvec(0,kk-1); |
---|
226 | } |
---|
227 | else |
---|
228 | if(indices.n_elem < kk) |
---|
229 | { |
---|
230 | const uvec indices2 = find(abs(eigval) <= tol2); |
---|
231 | |
---|
232 | const uword N_extra = (std::min)( indices2.n_elem, (kk - indices.n_elem) ); |
---|
233 | |
---|
234 | if(N_extra > 0) { indices = join_cols(indices, indices2.subvec(0,N_extra-1)); } |
---|
235 | } |
---|
236 | |
---|
237 | const uvec sorted_indices = sort_index(eigval, "descend"); |
---|
238 | |
---|
239 | S = eigval.elem(sorted_indices); S *= A_max; |
---|
240 | |
---|
241 | if(calc_UV) |
---|
242 | { |
---|
243 | uvec U_row_indices(A.n_rows); for(uword i=0; i < A.n_rows; ++i) { U_row_indices[i] = i; } |
---|
244 | uvec V_row_indices(A.n_cols); for(uword i=0; i < A.n_cols; ++i) { V_row_indices[i] = i + A.n_rows; } |
---|
245 | |
---|
246 | U = Datum<T>::sqrt2 * eigvec(U_row_indices, sorted_indices); |
---|
247 | V = Datum<T>::sqrt2 * eigvec(V_row_indices, sorted_indices); |
---|
248 | } |
---|
249 | } |
---|
250 | |
---|
251 | arma_debug_warn( (S.n_elem < k), "svds(): warning: found fewer singular values than specified" ); |
---|
252 | |
---|
253 | return true; |
---|
254 | } |
---|
255 | |
---|
256 | |
---|
257 | |
---|
258 | //! find the k largest singular values and corresponding singular vectors of sparse matrix X |
---|
259 | template<typename T1> |
---|
260 | inline |
---|
261 | bool |
---|
262 | svds |
---|
263 | ( |
---|
264 | Mat<typename T1::elem_type>& U, |
---|
265 | Col<typename T1::pod_type >& S, |
---|
266 | Mat<typename T1::elem_type>& V, |
---|
267 | const SpBase<typename T1::elem_type,T1>& X, |
---|
268 | const uword k, |
---|
269 | const typename T1::pod_type tol = 0.0, |
---|
270 | const typename arma_real_or_cx_only<typename T1::elem_type>::result* junk = 0 |
---|
271 | ) |
---|
272 | { |
---|
273 | arma_extra_debug_sigprint(); |
---|
274 | arma_ignore(junk); |
---|
275 | |
---|
276 | const bool status = svds_helper(U, S, V, X.get_ref(), k, tol, true); |
---|
277 | |
---|
278 | if(status == false) |
---|
279 | { |
---|
280 | arma_bad("svds(): failed to converge", false); |
---|
281 | } |
---|
282 | |
---|
283 | return status; |
---|
284 | } |
---|
285 | |
---|
286 | |
---|
287 | |
---|
288 | //! find the k largest singular values of sparse matrix X |
---|
289 | template<typename T1> |
---|
290 | inline |
---|
291 | bool |
---|
292 | svds |
---|
293 | ( |
---|
294 | Col<typename T1::pod_type >& S, |
---|
295 | const SpBase<typename T1::elem_type,T1>& X, |
---|
296 | const uword k, |
---|
297 | const typename T1::pod_type tol = 0.0, |
---|
298 | const typename arma_real_or_cx_only<typename T1::elem_type>::result* junk = 0 |
---|
299 | ) |
---|
300 | { |
---|
301 | arma_extra_debug_sigprint(); |
---|
302 | arma_ignore(junk); |
---|
303 | |
---|
304 | Mat<typename T1::elem_type> U; |
---|
305 | Mat<typename T1::elem_type> V; |
---|
306 | |
---|
307 | const bool status = svds_helper(U, S, V, X.get_ref(), k, tol, false); |
---|
308 | |
---|
309 | if(status == false) |
---|
310 | { |
---|
311 | arma_bad("svds(): failed to converge", false); |
---|
312 | } |
---|
313 | |
---|
314 | return status; |
---|
315 | } |
---|
316 | |
---|
317 | |
---|
318 | |
---|
319 | //! find the k largest singular values of sparse matrix X |
---|
320 | template<typename T1> |
---|
321 | inline |
---|
322 | Col<typename T1::pod_type> |
---|
323 | svds |
---|
324 | ( |
---|
325 | const SpBase<typename T1::elem_type,T1>& X, |
---|
326 | const uword k, |
---|
327 | const typename T1::pod_type tol = 0.0, |
---|
328 | const typename arma_real_or_cx_only<typename T1::elem_type>::result* junk = 0 |
---|
329 | ) |
---|
330 | { |
---|
331 | arma_extra_debug_sigprint(); |
---|
332 | arma_ignore(junk); |
---|
333 | |
---|
334 | Col<typename T1::pod_type> S; |
---|
335 | |
---|
336 | Mat<typename T1::elem_type> U; |
---|
337 | Mat<typename T1::elem_type> V; |
---|
338 | |
---|
339 | const bool status = svds_helper(U, S, V, X.get_ref(), k, tol, false); |
---|
340 | |
---|
341 | if(status == false) |
---|
342 | { |
---|
343 | arma_bad("svds(): failed to converge", true); |
---|
344 | } |
---|
345 | |
---|
346 | return S; |
---|
347 | } |
---|
348 | |
---|
349 | |
---|
350 | |
---|
351 | //! @} |
---|