Helper tool to create a statsfun for the options structure of solvers. function statsfun = statsfunhelper(name, fun) function statsfun = statsfunhelper(S) Usage with (name, fun): Input 1: name is a string which is a valid field name (no spaces, starts with a letter or an underscore, only alphanumeric characters and underscores). Input2: fun is a function handle with one output and 1 to 4 inputs, as follows (your choice): fun(x) or fun(problem, x) or fun(problem, x, stats) or fun(problem, x, stats, store) where the inputs are the ones that would be given to options.statsfun, as described in the help of the solver used. Typically, x is the point on the manifold at the current iterate, problem is the Manopt problem structure, stats is all the current statistics recorded for that iterate and store is the cache structure at the current iterate. When calling a Manopt solver with the options structure, such as for example with: [x, xcost, info] = steepestdescent(problem, [], options); you may set a field of the options structure as follows: options.statsfun = statsfunhelper('nameofthefield', fun); As a result, at each iteration, the stats structure will contain a field stats.nameofthefield with the value returned by the call to fun at that iterate. The stats structures are stored in the struct-array info. As an example, if the value returned by fun is a scalar, then [info.nameofthefield] is a vector containing all returned values. Usage with S: The input S is a structure. For each field of S, say S.field, the stats structure will be augmented with stats.field = fun(..), where fun is the function handle stored in S.field, and with the same conventions as above. This version allows to record more than one bit of information at each iteration. Example: metrics.nameofthefield = fun; metrics.othername = otherfun; options.statsfun = statsfunhelper(metrics); The different function handles (here, fun and otherfun) can take 1 to 4 inputs too, and they do not have to take the same number of inputs.

- PCA_stochastic Example of stochastic gradient algorithm in Manopt on a PCA problem.
- maxcut Algorithm to (try to) compute a maximum cut of a graph, via SDP approach.

0001 function statsfun = statsfunhelper(inp1, inp2) 0002 % Helper tool to create a statsfun for the options structure of solvers. 0003 % 0004 % function statsfun = statsfunhelper(name, fun) 0005 % function statsfun = statsfunhelper(S) 0006 % 0007 % Usage with (name, fun): 0008 % 0009 % Input 1: name is a string which is a valid field name (no spaces, starts 0010 % with a letter or an underscore, only alphanumeric characters and 0011 % underscores). 0012 % 0013 % Input2: fun is a function handle with one output and 1 to 4 inputs, as 0014 % follows (your choice): 0015 % 0016 % fun(x) or fun(problem, x) or 0017 % fun(problem, x, stats) or fun(problem, x, stats, store) 0018 % 0019 % where the inputs are the ones that would be given to options.statsfun, as 0020 % described in the help of the solver used. Typically, x is the point on 0021 % the manifold at the current iterate, problem is the Manopt problem 0022 % structure, stats is all the current statistics recorded for that iterate 0023 % and store is the cache structure at the current iterate. 0024 % 0025 % When calling a Manopt solver with the options structure, such as for 0026 % example with: 0027 % 0028 % [x, xcost, info] = steepestdescent(problem, [], options); 0029 % 0030 % you may set a field of the options structure as follows: 0031 % 0032 % options.statsfun = statsfunhelper('nameofthefield', fun); 0033 % 0034 % As a result, at each iteration, the stats structure will contain a field 0035 % stats.nameofthefield with the value returned by the call to fun at that 0036 % iterate. The stats structures are stored in the struct-array info. 0037 % As an example, if the value returned by fun is a scalar, then 0038 % [info.nameofthefield] is a vector containing all returned values. 0039 % 0040 % 0041 % Usage with S: 0042 % 0043 % The input S is a structure. For each field of S, say S.field, the stats 0044 % structure will be augmented with stats.field = fun(..), where fun is the 0045 % function handle stored in S.field, and with the same conventions as 0046 % above. This version allows to record more than one bit of information at 0047 % each iteration. Example: 0048 % 0049 % metrics.nameofthefield = fun; 0050 % metrics.othername = otherfun; 0051 % options.statsfun = statsfunhelper(metrics); 0052 % 0053 % The different function handles (here, fun and otherfun) can take 1 to 4 0054 % inputs too, and they do not have to take the same number of inputs. 0055 0056 % This file is part of Manopt: www.manopt.org. 0057 % Original author: Nicolas Boumal, Dec. 17, 2014. 0058 % Contributors: 0059 % Change log: 0060 0061 if (nargin == 1) && isstruct(inp1) 0062 S = inp1; 0063 elseif (nargin == 2) 0064 S = struct(inp1, inp2); 0065 else 0066 error('statsfunhelper takes 1 or 2 inputs. If 1 input, it must be a structure.'); 0067 end 0068 0069 0070 function stats = thestatsfun(problem, x, stats, store) 0071 names = fieldnames(S); 0072 for it = 1 : length(names) 0073 name = names{it}; 0074 fun = S.(name); 0075 switch nargin(fun) 0076 case 1 0077 stats.(name) = fun(x); 0078 case 2 0079 stats.(name) = fun(problem, x); 0080 case 3 0081 stats.(name) = fun(problem, x, stats); 0082 case 4 0083 stats.(name) = fun(problem, x, stats, store); 0084 otherwise 0085 error('The functions passed to statsfunhelper must take 1 to 4 inputs.'); 0086 end 0087 end 0088 end 0089 0090 statsfun = @thestatsfun; 0091 0092 end

Generated on Fri 08-Sep-2017 12:43:19 by