-
Notifications
You must be signed in to change notification settings - Fork 10
/
jmf_tests.m
414 lines (313 loc) · 12.1 KB
/
jmf_tests.m
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
406
407
408
409
410
411
412
413
414
function [] = jmf_tests(in1, in2)
%profile_run_l1();
%return
%if nargin >= 1
% if nargin == 1
% plot_time_scaling_driver_l1(in1);
% elseif nargin == 2
% plot_time_scaling_driver_l1(in1, in2);
% end
% return
%end
%run_time_scaling_l1();
%profile_run_l0();
%return
if nargin >= 1
plot_time_scaling_l0(in1);
return
end
run_time_scaling_l0();
end
%% SSC-l1
function [] = profile_run_l1()
profile off;
clear all; % profile clear doesn't clear all the way or something... this seems to do it though.
n = 600;
p = 256;
maxIter = 30;
X = randn(p,n);
affine = true
lambdaE = 1; % bogus parameter
shrink_mex2(struct('num_threads', 8));
prox_l1_and_sum_worker_mex(struct('num_threads', 8));
profile clear;
profile on;
% ADMM
% ---
%alpha_rho = 10;
%SSC_viaADMM(X, ...
% 'lambda', lambdaE, 'maxIter', maxIter, 'printEvery', 1,...
% 'tol', 1e-12, 'affine', affine, 'alpha_rho', alpha_rho, 'adaptiveRho', false,...
% 'errHistEvery', 1000, 'residHistEvery', 1000);
% TFOCS
% ---
%tfocs_opts = struct( 'errFcn', @(objective,C) errFcn(reshape(C,n,n)) );
tfocs_opts = struct();
SSC_viaProxGradient(X, 'lambda', lambdaE,...
'affine', affine, 'tfocs_opts', tfocs_opts , 'tol', 1e-6, 'maxIter', maxIter, ...
'printEvery', 1);
profile off
end
function [] = run_time_scaling_l1()
% Parameters for Section 4.2 of AAAI submission
% (comes from Stephen's paper_deom_3_cont.m)
% ---
rng(271828);
p = 256; % ambient dimension
K = 10; % number of subspaces
d = 3; % dimension of each subspace
sigma = 0.1; % noise standard deviation
affine = true % What we really want
%affine = false % less interesting, but easier prox to start speeding up
rho_all = 200:50:500; % the rho used in the paper
rho_all = round(linspace(20,50,5)); % bogus rho for quicker testing
alpha = [30,90];
maxIter = 30;
numExp = 5; % number of trials to run
shrink_mex2(struct('num_threads', 4)); % for linear and affine constraints, 8 threads doesn't help (doesn't hurt much either)
% this prox is memory bound?
prox_l1_and_sum_worker_mex(struct('num_threads', 8)); % cranking up the threads of this guy does help
% use no more that 8 threads (on my 16 core workstation)
% so that we stay fair to ADMM & other matlab stuff (e.g. matmat)
% which uses 8 threads.
ADMM_times = zeros(numel(rho_all), numExp);
TFOCS_times = zeros(numel(rho_all), numExp);
for densVal = 1 : numel(rho_all)
rho = rho_all(densVal);
N = rho * K * d
for trial = 1 : numExp
TmpBasis = randn(p, p);
[TmpBasis, ~] = qr(TmpBasis, 0);
X = zeros(p,N);
true_labels = zeros(N,1);
for i = 1 : K % for each subspace
Ni = rho *d;
ind = randperm(p,d);
U = TmpBasis(:,ind);
X(:, (i-1)*Ni+1:i*Ni) = U * randn(d, Ni) + sigma*randn(p,Ni);
true_labels((i-1)*Ni+1:i*Ni) = i;
end
% Set up solver params
% ---
threshold = 1e-6;
%lambdaE = []; %JMF 17 Nov 2018: if we don't specify lambda, the solver drivers will compute a lambda for us
% ADMM
alpha_rho = 1e-2; % TODO JMF 16 Nov 2018: this is a bogus parameter; okay for timing though?
adaptive_rho = true;
t_ = tic();
SSC_viaADMM(X, ...
'maxIter', maxIter, 'printEvery', 1,...
'tol', 1e-12, 'affine', affine, 'alpha_rho', alpha_rho, 'adaptiveRho', adaptive_rho,...
'errHistEvery', maxIter, 'residHistEvery', maxIter);
ADMM_times(densVal,trial) = toc(t_);
% TFOCS
%tfocs_opts = struct( 'errFcn', @(objective,C) errFcn(reshape(C,n,n)) );
tfocs_opts = struct();
t_ = tic();
SSC_viaProxGradient(X, ...
'affine', affine, 'tfocs_opts', tfocs_opts , 'tol', threshold, 'maxIter', maxIter, ...
'printEvery', 1);
TFOCS_times(densVal,trial) = toc(t_);
end
end
save('time_scaling_data.mat');
end
function [] = plot_time_scaling_driver_l1(in1, in2)
if nargin == 1
data_file = string(in1);
plot_time_scaling_single(data_file);
elseif nargin == 2
data_file_baseline = string(in1);
data_file_new = string(in2);
plot_time_scaling_compare(data_file_baseline, data_file_new);
else
error('unhandled number of input arguments');
end
end
function [] = plot_time_scaling_single_l1(data_file)
load(data_file);
ADMM_times_mean = mean(ADMM_times, 2);
TFOCS_times_mean = mean(TFOCS_times, 2);
ADMM_times_min = min(ADMM_times, [], 2);
TFOCS_times_min = min(TFOCS_times, [], 2);
ADMM_times_max = max(ADMM_times, [], 2);
TFOCS_times_max = max(TFOCS_times, [], 2);
% Plot runtime scaling on log-log
% ---
N_vec = rho_all * K * d;
figure(1);
clf;
hold on;
%plot(N_vec, ADMM_times_mean, 'LineWidth', 2);
%plot(N_vec, TFOCS_times_mean, 'LineWidth', 2);
errorbar(N_vec, ADMM_times_mean,...
ADMM_times_mean - ADMM_times_min, ADMM_times_max - ADMM_times_mean,...
'LineWidth', 2);
errorbar(N_vec, TFOCS_times_mean,...
TFOCS_times_mean - TFOCS_times_min, TFOCS_times_max - TFOCS_times_mean,...
'LineWidth', 2);
hold off;
set(gca, 'xscale', 'log');
set(gca, 'yscale', 'log');
xlim([N_vec(1) N_vec(end)]);
xlabel('number of data points');
ylabel('running time (sec.)');
if adaptive_rho
ADMM_label = 'AADMM';
else
ADMM_label = 'ADMM';
end
legend(ADMM_label, 'TFOCS', 'Location', 'NorthWest');
end
function [] = plot_time_scaling_compare_l1(data_file_baseline, data_file_new)
S_baseline = load(data_file_baseline);
S_new = load(data_file_new);
% Some sanity checks
% ---
if ~all(S_baseline.rho_all == S_new.rho_all)
error('rho_all vector doesn''t match in both files');
end
if S_baseline.affine ~= S_new.affine
error('linear/affine constraint doesn''t match in both files');
end
ADMM_times_mean_baseline = mean(S_baseline.ADMM_times, 2);
TFOCS_times_mean_baseline = mean(S_baseline.TFOCS_times, 2);
ADMM_times_min_baseline = min(S_baseline.ADMM_times, [], 2);
TFOCS_times_min_baseline = min(S_baseline.TFOCS_times, [], 2);
ADMM_times_max_baseline = max(S_baseline.ADMM_times, [], 2);
TFOCS_times_max_baseline = max(S_baseline.TFOCS_times, [], 2);
ADMM_times_mean_new = mean(S_new.ADMM_times, 2);
TFOCS_times_mean_new = mean(S_new.TFOCS_times, 2);
ADMM_times_min_new = min(S_new.ADMM_times, [], 2);
TFOCS_times_min_new = min(S_new.TFOCS_times, [], 2);
ADMM_times_max_new = max(S_new.ADMM_times, [], 2);
TFOCS_times_max_new = max(S_new.TFOCS_times, [], 2);
% Plot runtime scaling on log-log
% ---
N_vec = S_baseline.rho_all * S_baseline.K * S_baseline.d;
figure(1);
clf;
hold on;
%plot(N_vec, ADMM_times_mean_baseline, '-', 'LineWidth', 2);
%plot(N_vec, TFOCS_times_mean_baseline, '-', 'LineWidth', 2);
%plot(N_vec, ADMM_times_mean_new, '--', 'LineWidth', 2);
%plot(N_vec, TFOCS_times_mean_new, '--', 'LineWidth', 2);
errorbar(N_vec, ADMM_times_mean_baseline,...
ADMM_times_mean_baseline - ADMM_times_min_baseline, ADMM_times_max_baseline - ADMM_times_mean_baseline,...
'-', 'LineWidth', 2);
errorbar(N_vec, TFOCS_times_mean_baseline,...
TFOCS_times_mean_baseline - TFOCS_times_min_baseline, TFOCS_times_max_baseline - TFOCS_times_mean_baseline,...
'-', 'LineWidth', 2);
errorbar(N_vec, ADMM_times_mean_new,...
ADMM_times_mean_new - ADMM_times_min_new, ADMM_times_max_new - ADMM_times_mean_new,...
'--', 'LineWidth', 2);
errorbar(N_vec, TFOCS_times_mean_new,...
TFOCS_times_mean_new - TFOCS_times_min_new, TFOCS_times_max_new - TFOCS_times_mean_new,...
'--', 'LineWidth', 2);
hold off;
set(gca, 'xscale', 'log');
set(gca, 'yscale', 'log');
xlim([N_vec(1) N_vec(end)]);
xlabel('number of data points');
ylabel('running time (sec.)');
if S_baseline.adaptive_rho
ADMM_label_baseline = 'AADMM (baseline)';
else
ADMM_label_baseline = 'ADMM (baseline)';
end
if S_new.adaptive_rho
ADMM_label_new = 'AADMM (new)';
else
ADMM_label_new = 'ADMM (new)';
end
legend(ADMM_label_baseline, 'TFOCS (baseline)', ADMM_label_new, 'TFOCS (new)', 'Location', 'NorthWest');
end
%% SSC-l0
function [] = profile_run_l0()
profile off;
clear all; % profile clear doesn't clear all the way or something... this seems to do it though.
n = 3000;
p = 64;
k = 10;
maxIter = 30;
X = randn(p,n);
%inds = find(randn(p,n) > 2);
%X = sparse(p,n);
%X(inds) = randn(numel(inds),1);
%nnz(X)/prod(size(X))
affine = false
profile clear;
profile on;
% SSC-l0 via nonconvex proximal gradient method
% ---
driver = @SSC_viaNonconvexProxGradient;
driver = @SSC_viaNonconvexProxGradient2;
proj_largest_k_mex(struct('num_threads', 8));
proj_largest_k_affine_mex(struct('num_threads', 8));
driver(X, k, 'affine', affine, ...
'tol', -1e-6, 'maxIter', maxIter, 'printEvery', 1);
profile off
end
function [] = run_time_scaling_l0()
rng(271828);
p = 64;
n_vec = round(logspace(2,4,10));
k = 10;
affine = true;
%affine = false;
maxIter = 30;
numExp = 5; % number of trials to run
proj_largest_k_mex(struct('num_threads', 8));
proj_largest_k_affine_mex(struct('num_threads', 8));
orig_times = zeros(numel(n_vec), numExp);
new_times = zeros(numel(n_vec), numExp);
for i=1:numel(n_vec);
n = n_vec(i);
X = randn(p,n);
%inds = find(randn(p,n) > 2);
%X = sparse(p,n);
%X(inds) = randn(numel(inds),1);
%nnz(X)/prod(size(X))
for trial=1:numExp
t_ = tic();
SSC_viaNonconvexProxGradient(X, k, ...
'affine', affine, 'tol', -1e-6, 'maxIter', maxIter, 'printEvery', 1);
orig_times(i,trial) = toc(t_);
t_ = tic();
SSC_viaNonconvexProxGradient2(X, k, ...
'affine', affine, 'tol', -1e-6, 'maxIter', maxIter, 'printEvery', 1);
new_times(i,trial) = toc(t_);
end
end
save('time_scaling_data.mat');
end
function [] = plot_time_scaling_l0(data_file)
load(data_file)
orig_times_mean = mean(orig_times, 2);
new_times_mean = mean(new_times, 2);
orig_times_max = max(orig_times, [], 2);
new_times_max = max(new_times, [], 2);
orig_times_min = min(orig_times, [], 2);
new_times_min = min(new_times, [], 2);
% Plot runtime scaling on log-log
% ---
figure(1);
clf;
hold on;
errorbar(n_vec, orig_times_mean,...
orig_times_mean - orig_times_min, orig_times_max - orig_times_mean,...
'LineWidth', 2);
errorbar(n_vec, new_times_mean,...
new_times_mean - new_times_min, new_times_max - new_times_mean,...
'LineWidth', 2);
plot(n_vec, new_times_mean(1)*(n_vec/n_vec(1)), 'k--');
plot(n_vec, new_times_mean(1)*(n_vec/n_vec(1)).^2, 'k--');
hold off;
set(gca, 'xscale', 'log');
set(gca, 'yscale', 'log');
xlim([n_vec(1) n_vec(end)]);
xlabel('number of data points');
ylabel('running time (sec.)');
legend('\\ell\_0 - Original', '\\ell\_0 - New', 'Location', 'NorthWest');
title(sprintf('SSC \\ell_0 affine=%d', affine));
end