function [model,mn,st] = model_train(trnset,mn,st,ngmm,nsamp,dims)
% [model,mn,st] = model_train(trnset,mn,st,ngmm,nsamp,dims)
%   trnset is a list (cell array) of data files defining a class
%   mn, st are normalization constants (empty -> derive from data)
%   ngmm is the number of gaussian components (default 20), 
%   nsamp is the number of randomly-selcted frames to train on.
%   dims is a vector indexing which dimensions of the data to use (empty->all)
%   model is the model we return with fields mean, sigma, prior, nmix
%   mn, st are returned (e.g. to pass to next invocation)
% 2007-04-03 Dan Ellis dpwe@ee.columbia.edu  artist ID demo system

% Standardization parameters are taken from data if not specified
if nargin < 2;  mn = []; end
if nargin < 3;  st = []; end
% How many gaussian components?
if nargin < 4;  ngmm = 20; end
% Randomly sample how many data points?
if nargin < 5;  nsamp = 2000; end
% Which dimensions to use in the data?  Default to all
if nargin < 6;  dims = []; end

data = [];

guarddef = 1000;  % default guard spacing from ends

% Load all the data
for i = 1:length(trnset)
  d = readdatafile(trnset{i})';
  % only add the middle part of each file into the pool
  % ri will consist of ntsamp random indices (with replacement) at least 
  % <guard> samples away from end, where guard is larger of 1000 or one
  % quarter of the dataset length
  guard = min(round(size(d,1)/4),guarddef); % how far to stay away from ends
  ri = guard + ceil((size(d,1)-2*guard)*rand(1,nsamp));
  
  data = [data; d(ri,:)];
end

% Select random subset of frames
ri = randperm(size(data,1));
ri = ri(1:nsamp);

% If no dimensions explicitly selected, use them all
if length(dims) == 0
  dims = 1:length(mn);
end

% Select this data only
data = data(ri,:);

% Derive normalization constants if not specified
if length(mn) == 0
  [data,mn,st] = standardize(data');
  data = data';
else
  % Normalization by global constants (helps gmm??)
  data = standardize(data',mn,st)';
end

% remove unwanted dimensions now we've got all the stand coeffs
data = data(:,dims);

model = model_train_data(data, ngmm);