Added experiments with Oseledets TT-toolbox

eusipco/experiments_eusipco_tensortrain_amino.m

@@ -0,0 +1,76 @@
+%% Tensor Train Experiments
+
+addpath('./requiredfunctions/')
+addpath(strjoin(strcat('../thirdparty-matlab/oseledets-TT-Toolbox-8332a6c/',{'','core', 'exp', 'cross', 'fmex', 'misc','solve'}),';'))
+
+rng(131462234)
+%% Setup problem and generate data
+num_repeats = 10;
+maxiter = 50;
+
+load('../Data/Amino-Acid/claus.mat')
+N=size(X);
+
+
+% All permutations
+all_permuations = perms(1:ndims(X));
+all_permuations = all_permuations(end:-1:1,:); %Make sure the first one is 1:5
+num_permutes = size(all_permuations,1);
+
+%% Run experiments
+final_elbo = nan(num_permutes,num_repeats);
+final_rrmse = nan(num_permutes,num_repeats);
+final_tau = nan(num_permutes,num_repeats);
+final_tt_recon = nan(num_permutes,1);
+final_tt_rank = nan(num_permutes,ndims(X)+1);
+
+for i = 1:num_permutes
+    t0 = tic;
+    fprintf('Permutation %i of %i...', i, num_permutes)
+    perm_idx = all_permuations(i,:);
+
+    org_idx = 1:ndims(X);
+    perm2org_idx = sum(bsxfun(@times,(org_idx == perm_idx')', org_idx),2);
+
+    D_est_perm = [];
+
+    %% Calculate TT
+    tt = tt_tensor(permute(X,perm_idx), 1e-6);
+    cr= tt.core ; ps= tt.ps ;
+    for k=1:ndims(X)
+        core{k} =cr(ps(k): ps(k +1) -1); %#ok<SAGROW>
+        core{k} = reshape(core{k}, tt.r(k), tt.n(k), tt.r(k+1)); %#ok<SAGROW>
+    end
+    G_init = core;
+
+    X_recon = constructTensorTrain(G_init);
+    X_recon = permute(X_recon, perm2org_idx);
+    final_tt_recon(i) = norm(X(:)-X_recon(:),'fro')^2/norm(X(:),'fro')^2;
+    final_tt_rank(i,:) = tt.r;
+    final_numel_G = nan(num_permutes,1);
+    final_numel_G(i) = sum(cellfun(@numel, G_init));
+
+    %% Calculate prob from TT init
+    for j = 1:num_repeats
+
+        try
+            [G_est, S_est, V_est, tau_est, elbo] = tt_prob_tensor(permute(X,perm_idx), G_init, [],...
+                'maxiter',maxiter,'verbose','no', 'fixed_tau',0);
+
+            X_recon = constructTensorTrain(G_est);
+            X_recon = permute(X_recon, perm2org_idx);
+
+            final_rrmse(i,j) = norm(X(:)-X_recon(:),'fro')^2/norm(X(:),'fro')^2;
+            final_elbo(i,j) = elbo(end);
+            final_tau(i,j) = tau_est;
+        catch e
+            warning(sprintf('Something went wrong... Error message was:\n%s\n',e.message))
+        end
+    end
+
+    %%
+    %%
+    toc(t0)
+end
+save('./eusipco/synthetic_tensortrain_amino', 'final_elbo', 'final_rrmse', 'final_tau',...
+    'final_tt_recon', 'final_tt_rank','final_numel_G')

eusipco/experiments_eusipco_tensortrain_nonoise.m

@@ -0,0 +1,82 @@
+%% Tensor Train Experiments
+addpath('./requiredfunctions/')
+addpath(strjoin(strcat('../thirdparty-matlab/oseledets-TT-Toolbox-8332a6c/',{'','core', 'exp', 'cross', 'fmex', 'misc','solve'}),';'))
+
+rng(131462234)
+%% Setup problem and generate data
+num_repeats = 10;
+maxiter = 50;
+
+N = 20:-1:16;
+D = [1,length(N)+1:-1:3, 1];
+D_est = D;
+
+[X_clean, G, ES, EV] = generateTensorTrain(N, D);
+X_clean = X_clean/sqrt(var(X_clean(:)));
+
+% All permutations
+all_permuations = perms(1:ndims(X_clean));
+all_permuations = all_permuations(end:-1:1,:); %Make sure the first one is 1:5
+num_permutes = size(all_permuations,1);
+
+%%
+num_noise_level = length(list_snrdb);
+%% Run experiments
+final_elbo = nan(num_permutes,num_repeats);
+final_rrmse = nan(num_permutes,num_repeats);
+final_tau = nan(num_permutes,num_repeats);
+final_tt_recon = nan(num_permutes,1);
+final_tt_rank = nan(num_permutes,length(D));
+final_numel_G = nan(num_permutes,1);
+
+X=X_clean;%
+%X=addTensorNoise(X_clean, 100); % Added 5 dB noise
+
+for i = 1:num_permutes
+    t0 = tic;
+    fprintf('Permutation %i of %i...', i, num_permutes)
+    perm_idx = all_permuations(i,:);
+
+    org_idx = 1:ndims(X);
+    perm2org_idx = sum(bsxfun(@times,(org_idx == perm_idx')', org_idx),2);
+
+    D_est_perm = [];
+
+    %% Calculate TT
+    tt = tt_tensor(permute(X,perm_idx), 1e-6);
+    cr= tt.core ; ps= tt.ps ;
+    for k=1:5
+        core{k} =cr(ps(k): ps(k +1) -1); %#ok<SAGROW>
+        core{k} = reshape(core{k}, tt.r(k), tt.n(k), tt.r(k+1)); %#ok<SAGROW>
+    end
+    G_init = core;
+
+    X_recon = constructTensorTrain(G_init);
+    X_recon = permute(X_recon, perm2org_idx);
+    final_tt_recon(i) = norm(X(:)-X_recon(:),'fro')^2/norm(X(:),'fro')^2;
+    final_tt_rank(i,:) = tt.r;
+    final_numel_G(i) = sum(cellfun(@numel, G_init));
+    %% Calculate prob from TT init
+    for j = 1:num_repeats
+
+        try
+            [G_est, S_est, V_est, tau_est, elbo] = tt_prob_tensor(permute(X,perm_idx), G_init, [],...
+                'maxiter',maxiter,'verbose','no', 'fixed_tau',0);
+
+            X_recon = constructTensorTrain(G_est);
+            X_recon = permute(X_recon, perm2org_idx);
+
+            final_rrmse(i,j) = norm(X(:)-X_recon(:),'fro')^2/norm(X(:),'fro')^2;
+            final_elbo(i,j) = elbo(end);
+            final_tau(i,j) = tau_est;
+        catch e
+            warning(sprintf('Something went wrong... Error message was:\n%s\n',e.message))
+        end
+    end
+
+    %%
+    %%
+    toc(t0)
+end
+save('./eusipco/synthetic_tensortrain_cleandata', 'final_elbo', 'final_rrmse', 'final_tau',...
+    'final_tt_recon', 'final_tt_rank','final_numel_G')

run_experiments_eusipco.m

@@ -0,0 +1,24 @@
+%% Run Experiments for EUSIPCO
+addpath(genpath('./'))
+addpath(genpath('../ncptensor'))
+addpath(genpath('../tools'))
+
+elapsed_time = zeros(6,1)*nan;
+for i = 3:6
+    t0 = tic;
+    switch i
+        case 1
+            experiments_aminoacid
+        case 2
+            experiments_aminoacid_modelorder
+        case 3
+            experiments_eusipco_tensortrain_nonoise
+        case 4
+            experiments_eusipco_tensortrain_amino
+        case 5
+            experiments_eusipco_knownD
+        case 6
+            experiments_eusipco_unknownD
+    end
+    elapsed_time(i) = toc(t0);
+end