Mercurial > hg > camir-aes2014
view toolboxes/MIRtoolbox1.3.2/somtoolbox/som_vs1to2.m @ 0:e9a9cd732c1e tip
first hg version after svn
author | wolffd |
---|---|
date | Tue, 10 Feb 2015 15:05:51 +0000 |
parents | |
children |
line wrap: on
line source
function sS = som_vs1to2(sS) %SOM_VS1TO2 Convert version 1 structure to version 2. % % sSnew = som_vs1to2(sSold) % % sMnew = som_vs1to2(sMold); % sDnew = som_vs1to2(sDold); % % Input and output arguments: % sSold (struct) a SOM Toolbox version 1 structure % sSnew (struct) a SOM Toolbox version 2 structure % % For more help, try 'type som_vs1to2' or check out online documentation. % See also SOM_SET, SOM_VS2TO1. %%%%%%%%%%%%% DETAILED DESCRIPTION %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% % % som_vs1to2 % % PURPOSE % % Transforms SOM Toolbox 1 version structs from to 2 version structs. % % SYNTAX % % sS2 = som_vs1to2(sS1) % % DESCRIPTION % % This function is offered to allow the change of old map and data structs % to new ones. There are quite a lot of changes between the versions, % especially in the map struct, and this function makes it easy to update % the structs. % % WARNING! % % 'som_unit_norm' normalization type is not supported by version 2, % so this type of normalization will be lost. % % REQUIRED INPUT ARGUMENTS % % sS1 (struct) any SOM Toolbox version 1 struct (map, data, % training or normalization struct) % % OUTPUT ARGUMENTS % % sS2 (struct) the corresponding SOM Toolbox 2 version struct % % EXAMPLES % % sM = som_vs1to2(sMold); % sD = som_vs1to2(sDold); % sT = som_vs1to2(sMold.train_sequence{1}); % sN = som_vs1to2(sDold.normalization); % % SEE ALSO % % som_set Set values and create SOM Toolbox structs. % som_vs2to1 Transform structs from version 2.0 to 1.0. % Copyright (c) 1999-2000 by the SOM toolbox programming team. % http://www.cis.hut.fi/projects/somtoolbox/ % Version 2.0beta juuso 101199 %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% %% check arguments error(nargchk(1, 1, nargin)); % check no. of input arguments is correct %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% %% set field values if isfield(sS,'codebook'), type='som_map'; elseif isfield(sS,'data'), type='som_data'; elseif isfield(sS,'algorithm'), type = 'som_train'; elseif isfield(sS,'inv_params'), type = 'som_norm'; else error('Unrecognized input struct.'); end switch type, case 'som_map', msize = sS.msize; munits = prod(msize); dim = prod(size(sS.codebook))/munits; M = reshape(sS.codebook,[munits dim]); % topology if strcmp(sS.shape,'rect'), shape = 'sheet'; else shape = sS.shape; end sTopol = struct('type','som_topol','msize',msize,'lattice',sS.lattice,'shape',shape); % labels labels = cell(munits,1); for i=1:munits, for j=1:length(sS.labels{i}), labels{i,j} = sS.labels{i}{j}; end end % trainhist tl = length(sS.train_sequence); if strcmp(sS.init_type,'linear'); alg = 'lininit'; else alg = 'randinit'; end trh = struct('type','som_train'); trh.algorithm = alg; trh.neigh = sS.neigh; trh.mask = sS.mask; trh.data_name = sS.data_name; trh.radius_ini = NaN; trh.radius_fin = NaN; trh.alpha_ini = NaN; trh.alpha_type = ''; trh.trainlen = NaN; trh.time = ''; for i=1:tl, trh(i+1) = som_vs1to2(sS.train_sequence{i}); trh(i+1).mask = sS.mask; trh(i+1).neigh = sS.neigh; trh(i+1).data_name = sS.data_name; end % component normalizations cnorm = som_vs1to2(sS.normalization); if isempty(cnorm), cnorm = cell(dim,1); elseif length(cnorm) ~= dim, warning('Incorrect number of normalizations. Normalizations ignored.\n'); cnorm = cell(dim,1); else if strcmp(cnorm{1}.method,'histD'), M = redo_hist_norm(M,sS.normalization.inv_params,cnorm); end end % map sSnew = struct('type','som_map'); sSnew.codebook = M; sSnew.topol = sTopol; sSnew.labels = labels; sSnew.neigh = sS.neigh; sSnew.mask = sS.mask; sSnew.trainhist = trh; sSnew.name = sS.name; sSnew.comp_norm = cnorm; sSnew.comp_names = sS.comp_names; case 'som_data', [dlen dim] = size(sS.data); % component normalizations cnorm = som_vs1to2(sS.normalization); if isempty(cnorm), cnorm = cell(dim,1); elseif length(cnorm) ~= dim, warning('Incorrect number of normalizations. Normalizations ignored.\n'); cnorm = cell(dim,1); else if strcmp(cnorm{1}.method,'histD'), sS.data = redo_hist_norm(sS.data,sS.normalization.inv_params,cnorm); end end % data sSnew = struct('type','som_data'); sSnew.data = sS.data; sSnew.name = sS.name; sSnew.labels = sS.labels; sSnew.comp_names = sS.comp_names; sSnew.comp_norm = cnorm; sSnew.label_names = []; case 'som_norm', if isempty(sS.inv_params), sSnew = []; else dim = size(sS.inv_params,2); sSnew = cell(dim,1); switch sS.name, case 'som_var_norm', method = 'var'; case 'som_lin_norm', method = 'range'; case 'som_hist_norm', method = 'histD'; case 'som_unit_norm', method = ''; warning(['Normalization method ''som_unit_norm'' is not available' ... ' in version 2 of SOM Toolbox.\n']); end if ~isempty(method), for i=1:dim, sSnew{i} = struct('type','som_norm'); sSnew{i}.method = method; sSnew{i}.params = []; sSnew{i}.status = 'done'; switch method, case 'var', me = sS.inv_params(1,i); st = sS.inv_params(2,i); sSnew{i}.params = [me, st]; case 'range', mi = sS.inv_params(1,i); ma = sS.inv_params(2,i); sSnew{i}.params = [mi, ma-mi]; case 'histD', vals = sS.inv_params(1:(end-1),i); bins = sum(isfinite(vals)); vals = vals(1:bins); sSnew{i}.params = vals; end end end end case 'som_train', sSnew = struct('type','som_train'); sSnew.algorithm = sS.algorithm; sSnew.neigh = 'gaussian'; sSnew.mask = []; sSnew.data_name = 'unknown'; sSnew.radius_ini = sS.radius_ini; sSnew.radius_fin = sS.radius_fin; sSnew.alpha_ini = sS.alpha_ini; sSnew.alpha_type = sS.alpha_type; sSnew.trainlen = sS.trainlen; sSnew.time = sS.time; case 'som_topol', disp('Version 1.0 of SOM Toolbox did not have topology structure.\n'); case {'som_grid','som_vis'} disp('Version 1.0 of SOM Toolbox did not have visualization structs.\n'); otherwise, error('Unrecognized struct.'); end sS = sSnew; %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% %% subfunctions function D = redo_hist_norm(D,inv_params,cnorm) dim = size(D,2); % first - undo the old way n_bins = inv_params(end,:); D = round(D * sparse(diag(n_bins))); for i = 1:dim, if any(isnan(D(:, i))), D(isnan(D(:, i)), i) = n_bins(i); end D(:, i) = inv_params(D(:, i), i); end % then - redo the new way for i=1:dim, bins = length(cnorm{i}.params); x = D(:,i); inds = find(~isnan(x) & ~isinf(x))'; for j = inds, [dummy ind] = min(abs(x(j) - cnorm{i}.params)); if x(j) > cnorm{i}.params(ind) & ind < bins, x(j) = ind + 1; else x(j) = ind; end end D(:,i) = (x-1)/(bins-1); end