Version 1.0.2

Author: Hiroyuki KASAI

README.md

@@ -1,10 +1,11 @@
 # RSOpt (Riemannian stochastic optimization algorithms)
 
 Authors: [Hiroyuki Kasai](http://www.kasailab.com/kasai_e.htm)
+Collaborators: [Bamdev Mishra](https://bamdevmishra.in/) and [Hiroyuki Sato](https://sites.google.com/site/hiroyukisatoeng/)
 
-Last page update: July 20, 2018
+Last page update: September 13, 2018
 
-Latest version: 1.0.1 (see Release notes for more info) 
+Latest version: 1.0.2 (see Release notes for more info) 
 
 <br />
 
@@ -132,6 +133,8 @@ If you have any problems or questions, please contact the author: [Hiroyuki Kasa
 
 Release Notes
 --------------
+* Version 1.0.2 (Sep. 13, 2018)
+    - MC problem (with Jester dataset) example is added. 
 * Version 1.0.1 (July 20, 2018)
     - Initial codes are available.
 * Version 1.0.0 (July 12, 2018)

dataset/jester/jester_mat.mat

@@ -1,15 +1,18 @@
 function demo()
+% This file is part of RSOpt package.
+%
+% Created by H.Kasai and B.Mishra on July 20, 2018
 
     clc; close all; clear
 
 
-    %% Define parameters
+    %% define parameters
     N = 500;
     d = 3;    
 
 
 
-    %% Read dataset
+    %% read dataset
     input_data = load('./dataset/psd/psd_mean_3_500_5.mat');
     A = zeros(d, d, N);
     A = input_data.x_sample{1};
@@ -18,13 +21,13 @@ function demo()
 
 
 
-    %% Set manifold
+    %% set manifold
     problem.M = sympositivedefinitefactory_mod(d);  
     problem.ncostterms = N;     
 
 
-    
-    %% Cost function
+    %% define problem
+    % cost function
     problem.cost = @cost;    
     function f = cost(X)
         f=0;
@@ -46,7 +49,7 @@ function demo()
 
 
 
-    %% Riemannian gradient of the cost function
+    % Riemannian gradient of the cost function
     problem.rgrad = @rgrad;      
     function g = rgrad(X)
 
@@ -64,7 +67,7 @@ function demo()
 
 
 
-    %% Riemannian stochastic gradient of the cost function
+    % Riemannian stochastic gradient of the cost function
     problem.partialgrad = @partialgrad;
     function g = partialgrad(X, idx_batchsize)        
 
@@ -88,7 +91,7 @@ function demo()
 
 
 
-    %% Run SRG algorithms    
+    %% run SRG algorithm    
 
     Init = problem.M.rand();
 
@@ -107,7 +110,7 @@ function demo()
 
 
 
-    %% Plots
+    %% plots
     fs = 20;
 
     % Optimality gap (Train loss - optimum) versus #grads/N     

demo.m

@@ -1,8 +1,12 @@
 function show_centroid_plots()
+% This file is part of RSOpt package.
+%
+% Created by H.Kasai and B.Mishra on July 20, 2018
+% Modified by H.Kasai on Sep. 13, 2018
 
     clc; close all; clear
 
-    %% Define parameters
+    %% define parameters
     tolgradnorm = 1e-8;
     inner_repeat = 1;
 
@@ -22,28 +26,31 @@ function show_centroid_plots()
     cn = 5; 
     srg_varpi = 0.05;
 end
-    
 
+
+
+
+    %% generate dataset
     input_filename = sprintf('./dataset/psd/psd_mean_%d_%d_%d.mat', d, N, cn);
     fprintf('Reading file %s with (d:%d N:%d cn:%d) .... ', input_filename, d, N, cn); 
     input_data = load(input_filename);        
-
     A = zeros(d, d, N);
-
     A = input_data.x_sample{1};
     fprintf('done\n');
-
     f_sol = input_data.f_sol{1}; 
     fprintf('f_sol: %.16e\n', f_sol);           
 
 
-    %% Set manifold
+    
+    
+    
+    %% set manifold
     problem.M = sympositivedefinitefactory_mod(d);  
     problem.ncostterms = N;     
 
 
 
-    % Cost function
+    %% define problem
     problem.cost = @cost;    
     function f = cost(X)
         f=0;
@@ -60,11 +67,9 @@ function show_centroid_plots()
             f = f + norm(logm(arg),'fro')^2;
         end
 
-        f = f/(N);
+        f = f/(2*N);
     end
 
-
-
     % Riemannian gradient of the cost function
     problem.rgrad = @rgrad;      
     function g = rgrad(X)
@@ -78,11 +83,10 @@ function show_centroid_plots()
 
         g = 2*X*logsum;
         g = (g+g')/2;
-        g = g/N;
+        g = g/(2*N);
     end
 
 
-
     % Riemannian stochastic gradient of the cost function
     problem.partialgrad = @partialgrad;
     function g = partialgrad(X, idx_batchsize)        
@@ -106,27 +110,29 @@ function show_centroid_plots()
 
 
 
-    %% Run algorithms    
 
-    % Initialize
+    
+    %% run algorithms    
+    
+    % initialize
     Uinit = problem.M.rand();
 
 
-    % Run R-SD
+    % R-SD
     clear options;
     options.maxiter = maxepoch;
     options.tolgradnorm = tolgradnorm;         
     [~, ~, infos_sd, options_sd] = steepestdescent(problem, Uinit, options); 
 
 
-    % Run R-CG
+    % R-CG
     clear options;
     options.maxiter = maxepoch;
     options.tolgradnorm = tolgradnorm;         
     [~, ~, infos_cg, options_cg] = conjugategradient(problem, Uinit, options);     
 
 
-    % Run SGD with decay step-size
+    % R-SGD with decay step-size
     clear options;
     options.verbosity = 1;
     options.batchsize = 10;
@@ -141,7 +147,7 @@ function show_centroid_plots()
     [~, ~, infos_sgd, options_sgd] = Riemannian_svrg(problem, Uinit, options);
 
 
-    % Run SVRG
+    % R-SVRG
     clear options;
     options.verbosity = 1;
     options.batchsize = 10;
@@ -158,7 +164,7 @@ function show_centroid_plots()
     [~, ~, infos_svrg, options_svrg] = Riemannian_svrg(problem, Uinit, options);
 
 
-    % Run SRG
+    % R-SRG
     clear options;
     options.verbosity = 1;
     options.batchsize = 10;
@@ -187,7 +193,7 @@ function show_centroid_plots()
     [~, ~, infos_srg_plus, options_srg_plus] = Riemannian_srg(problem, Uinit, options);  
 
 
-    % Calculate # of gradient evaluations
+    % calculate # of gradient evaluations
     num_grads_sd = (1:length([infos_sd.cost])) - 1; % N*options_sd.maxiter;
     num_grads_cg = (1:length([infos_cg.cost])) - 1; % N*options_sd.maxiter;    
     num_grads_sgd = ceil(options_sgd.maxinneriter/N)*((1:length([infos_sgd.cost])) - 1); % options.maxepoch*(options_sgd.maxinneriter);
@@ -206,7 +212,7 @@ function show_centroid_plots()
 
 
 
-    %% Plots
+    %% plots
     fs = 20;
 
     % Optimality gap (Train loss - optimum) versus #grads/N