%----------------------------------------------------------------------
% standard mean squared error search
-function [idx contrib errors test_ g mg] = vq_search_mse(vq, data)
+function [idx contrib errors test_ g mg sl] = vq_search_mse(vq, data)
[nVec nCols] = size(vq);
nRows = rows(data);
test_(f,:) = vq(min_ind,:);
end
- g = mg = 1; % dummys for this function
+ g = 0; mg = 1; sl = 0; % dummys for this function
endfunction
%----------------------------------------------------------------------
% abs() search with a linear gain term
-function [idx contrib errors test_ g mg] = vq_search_gain(vq, data)
+function [idx contrib errors test_ g mg sl] = vq_search_gain(vq, data)
[nVec nCols] = size(vq);
nRows = rows(data);
errors(f) = mn;
contrib(f,:) = test_(f,:) = vq(min_ind,:) + g(f,min_ind);
end
- mg = 1;
+ mg = 1; sl = 0;
endfunction
%----------------------------------------------------------------------
% abs() search with a linear plus ampl scaling term
-function [idx contrib errors test_ g mg] = vq_search_mag(vq, data)
+function [idx contrib errors test_ g mg sl] = vq_search_mag(vq, data)
[nVec nCols] = size(vq);
nRows = rows(data);
contrib(f,:) = test_(f,:) = mg(f,min_ind) * vq(min_ind,:) + g(f,min_ind);
end
-
+ sl = 0;
endfunction
[nRows nCols] = size(test);
- [idx contrib errors test_ g mg] = feval(search_func, vq, test);
-
+ [idx contrib errors test_ g mg sl] = feval(search_func, vq, test);
+
% sd over time
sd_per_frame = zeros(nRows,1);
sd_per_frame(i) = mean(abs(test(i,:) - test_(i,:)));
end
+ printf("average sd: %3.2f\n", mean(sd_per_frame));
%printf("%18s nVec: %d SD: %3.2f dB\n", search_func, nVec, mean(sd_per_frame));
- % plots sd and errors over time
+ % Optional GUI with U to anlayse VQ perf ---------------------------------------------------
if gui_en
- figure(1); clf; subplot(211); plot(sd_per_frame); title('SD'); subplot(212); plot(errors); title('mean error');
+ gui_mode = "sort"; % "sort" or "fbf"
+ sort_order = "worst"; % "worst" or "best"
+ f = 1;
+
+ % plots sd over time and histogram
+
+ figure(1); clf; subplot(211); plot(sd_per_frame); title('SD');
+ subplot(212); [yy xx] = hist(sd_per_frame); plot(xx,yy,'+-');
+ end
+
+ while gui_en
% display m frames, printing some stats, plotting vector to give visual idea of match
figure(2); clf;
- [errors_dec frame_dec] = sort(errors, "descend");
- m = 4;
+
+ % try this in 'ascend' or 'descend' mode to best or worst frames
+
+ if strcmp(gui_mode, "sort")
+ if strcmp(sort_order, "best")
+ [errors_dec frame_dec] = sort(errors, "ascend");
+ end
+ if strcmp(sort_order, "worst")
+ [errors_dec frame_dec] = sort(errors, "descend");
+ end
+ m = 4;
+ else
+ m = 1;
+ errors_dec = errors(f); frame_dec(1) = f;
+ end
+
for i=1:m
+
+ % build up VQ legend
+
af = frame_dec(i); aind = idx(af);
- l = sprintf("idx: %d", aind);
- ag = 0; amg = 1;
- if strcmp(search_func, "vq_search_gain") || strcmp(search_func, "vq_search_mag")
+
+ ag = 0; amg = 1; asl = 0;
+ if strcmp(search_func, "vq_search_gain")
ag = g(af,aind);
- l = sprintf("%s g: %3.2f", l, ag);
+ l3 = sprintf("g: %3.2f", ag);
end
if strcmp(search_func, "vq_search_mag")
+ ag = g(af,aind);
+ amg = mg(af,aind);
+ l3 = sprintf("g: %3.2f mg: %3.2f", ag, amg);
+ end
+ if strcmp(search_func, "vq_search_slope")
+ ag = g(af,aind);
amg = mg(af,aind);
- l = sprintf("%s mg: %3.2f", l, amg);
+ asl = sl(af,aind);
+ l3 = sprintf("g: %3.2f mg: %3.2f sl: %3.2f", ag, amg, asl);
end
- %printf("%d f: %d %s\n", i, af, l);
- subplot(sqrt(m),sqrt(m),i);
- l1 = sprintf("b-;fr %d;", af);
+ % plot target
+
+ subplot(sqrt(m), sqrt(m),i);
+ l1 = sprintf("b-;fr %d sd: %3.2f;", af, errors_dec(m));
plot(test(af,:), l1);
+ % plot vq vector and modified version
+
hold on;
l2 = sprintf("g-+;ind %d;", aind);
plot(vq(aind, :), l2);
- l3 = sprintf("g-o;%s;",l);
- plot(amg*vq(aind, :) + ag, l3);
+ l3 = sprintf("g-o;%s;",l3);
+ plot(amg*vq(aind, :) + ag + asl*(1:nCols), l3);
hold off;
axis([1 nCols -10 40]);
end
+
+ % interactive menu ------------------------------------------
+
+ if strcmp(gui_mode, "sort")
+ printf("\rmenu: m-mode[%s] o-order[%s] q-quit", gui_mode, sort_order);
+ end
+ if strcmp(gui_mode, "fbf")
+ printf("\rmenu: m-mode[%s] frame: %d n-next b-back q-quit", gui_mode, f);
+ end
+ fflush(stdout);
+ k = kbhit();
+
+ if k == 'm'
+ if strcmp(gui_mode, "sort")
+ gui_mode = "fbf"; f = frame_dec(1);
+ else
+ gui_mode = "sort";
+ end;
+ end
+ if k == 'o'
+ if strcmp(sort_order, "worst")
+ sort_order = "best";
+ else
+ sort_order = "worst";
+ end;
+ end
+ if k == 'n', f = f + 1; endif;
+ if k == 'b', f = f - 1; endif;
+ if k == 'q', gui_en = 0; printf("\n"); endif;
end
endfunction
Nvec = sim_in.Nvec;
train_func = get_search_func(sim_in.train_func_short);
- [idx vq] = kmeans(sim_in.trainvec, Nvec,
- "start", "sample",
- "emptyaction", "singleton",
- "search_func", train_func);
+ printf(" Nvec %d kmeans start\n", Nvec);
+ [idx vq] = feval(sim_in.kmeans,
+ sim_in.trainvec, Nvec,
+ "start", "sample",
+ "emptyaction", "singleton",
+ "search_func", train_func);
+ printf(" Nvec %d kmeans end\n", Nvec);
sd = []; des = []; tests = sim_in.tests;
for t=1:length(cellstr(tests))
for j=1:length(cellstr(test_list))
test_func_short = char(cellstr(test_list)(j));
[y x] = search_tests(sd, desc, train_func_short, test_func_short);
- leg = sprintf("o-%d;train: %5s test: %5s;", nlines, train_func_short, test_func_short)
+ leg = sprintf("o-%d;train: %5s test: %5s;", nlines, train_func_short, test_func_short);
if nlines, hold on; end;
x += inc; inc += 0.1; % separate x coords a bit to make errors bars legible
errorbar(x, y(:,1), y(:,2), leg);
%----------------------------------------------------------------------
% Run a bunch of long tests in parallel and plot results
+%
+
+% This test tries a bunch of training and test search combinations on
+% the first 1kHz to see if there is any advantage in using the higher
+% order search techniques in the VQ training. So far it appears
+% not.... However the higher order search routines do give great
+% results compared to a basic MSE (or oder) search.
-function [sd desc] = long_tests(quick_check=0)
+function [sd desc] = long_tests_1_10(quick_check=0)
num_cores = 4;
K = 10;
load surf_train_120_hpf150; load surf_all_hpf150;
% build up a big array of tests to run ------------------------
- sim_in.trainvec = trainvec; sim_in.testvec = testvec;
+ sim_in.trainvec = trainvec; sim_in.testvec = testvec;
+ sim_in.kmeans = "kmeans2"; % slow but supports different search functions
sim_in.Nvec = 64;
sim_in.train_func_short = "mse";
sim_in.tests = ["mse"; "gain"; "mag"; "slope"];
% run test list in parallel
[sd desc] = pararrayfun(num_cores, @train_vq_and_run_tests, test_list);
-
+ %train_vq_and_run_tests(test_list(1));
+
% Plot results -----------------------------------------------
fg = 1;
endfunction
-function short_detailed_test(train_func, test_func)
- K = 10;
+%----------------------------------------------------------------------
+% This test tries some test combinations on the 1000 to 3000 Hz range.
+%
+% Based on our results with long_test_1_10() we just use MSE training
+
+function [sd desc] = long_tests_11_40(quick_check=0)
+ num_cores = 4;
+ K = 20; st = 11; en = 40;
+ load surf_train_120_hpf150; load surf_all_hpf150;
+
+ if quick_check
+ NtrainVec = 1000;
+ NtestVec = 100;
+ else
+ NtrainVec = length(surf_train_120_hpf150);
+ NtestVec = length(surf_all_hpf150);
+ end
+
+ trainvec = surf_train_120_hpf150(1:NtrainVec,st:en);
+ testvec = surf_all_hpf150(1:NtestVec,st:en);
+
+ % build up an array of tests to run ------------------------
+
+ sim_in.kmeans = "kmeans";
+ sim_in.trainvec = trainvec; sim_in.testvec = testvec;
+ sim_in.Nvec = 64;
+ sim_in.train_func_short = "mse";
+ sim_in.tests = ["slope"];
+
+ % Test1: mse training, 64, 128, 256, 512
+
+ sim_in_vec(1:4) = sim_in;
+ sim_in_vec(2).Nvec = 128; sim_in_vec(3).Nvec = 256; sim_in_vec(4).Nvec = 512;
+
+ test_list = sim_in_vec;
+
+ % run test list in parallel
+
+ [sd desc] = pararrayfun(num_cores, @train_vq_and_run_tests, test_list);
+ %[sd desc] = train_vq_and_run_tests(sim_in_vec(4));
+
+ % Plot results -----------------------------------------------
+
+ fg = 2;
+ plot_sd_results(fg++, "MSE Training 10..30", sd, desc, "mse", "slope");
+
+#{
+ % histogram of results from Nvec=64, all training methods, slope search
+
+ testnum = search_tests_Nvec(sd, desc, "mse", "slope", 64);
+ hist_sd = sd(:,testnum); hist_desc = desc(testnum);
+ testnum = search_tests_Nvec(sd, desc, "gain", "slope", 64);
+ hist_sd = [hist_sd sd(:,testnum)]; hist_desc = [hist_desc desc(testnum)];
+ testnum = search_tests_Nvec(sd, desc, "mag", "slope", 64);
+ hist_sd = [hist_sd sd(:,testnum)]; hist_desc = [hist_desc desc(testnum)];
+ testnum = search_tests_Nvec(sd, desc, "slope", "slope", 64);
+ hist_sd = [hist_sd sd(:,testnum)]; hist_desc = [hist_desc desc(testnum)];
+ compare_hist(fg, "Histogram of SDs Nvec=64", hist_sd, hist_desc);
+#}
+endfunction
+
+
+function vq = detailed_test(Nvec=64, st=1, en=10, quick_check=1, test_func = "vq_search_slope")
+ K = en-st+1;
load surf_train_120_hpf150;
- load surf_all;
- NtrainVec = 1000;
- NtestVec = 100;
- trainvec = surf_train_120_hpf150(1:NtrainVec,1:K);
- testvec = surf_all(1:NtestVec,1:K);
+ load surf_all_hpf150;
+ if quick_check
+ NtrainVec = 1000;
+ NtestVec = 100;
+ else
+ NtrainVec = length(surf_train_120_hpf150);
+ NtestVec = length(surf_all_hpf150);
+ NtestVec = 100;
+ end
- Nvec = 64; % we can plot all vectors on one screen of subplots
-
- [idx vq] = kmeans2(trainvec, Nvec,
- "start", "sample",
- "emptyaction", "singleton",
- "search_func", train_func);
+ trainvec = surf_train_120_hpf150(1:NtrainVec,st:en);
+ testvec = surf_all_hpf150(1:NtestVec,st:en);
- sdpf = run_test(vq, testvec, Nvec, test_func, gui_en=1);
+ [idx vq] = kmeans(trainvec, Nvec,
+ "start", "sample",
+ "emptyaction", "singleton");
+
+ run_test(vq, testvec, Nvec, test_func, gui_en=1);
endfunction
-% Some contrived examples to test VQ training
+% ------------------------------------------------------------------
+%
+% Following test_* functions have some contrived examples to test VQ
+% training with each training method. However the longtest_1_10
+% results suggest these training methds don't provide any improvement
+% over standard MSE kmeans.
function test_training_mse
K = 3; NtrainVec = 10; Nvec = 2;
rand('seed',1); % kmeans using rand for initial population,
% we want same results on every run
-%short_detailed_test('vq_search_mag', 'vq_search_mag');
-[sd desc] = long_tests(quick_check=0);
+% choose one of these to run
+
+detailed_test(64, 1, 20, quick_check=1);
+%[sd desc] = long_tests_1_10(quick_check=1);
+%[sd desc] = long_tests_11_40(quick_check=0);
%test_training_slope
projects / freetel-svn-tracking.git / commitdiff