nnet_demo_2.m

function nnet_demo_2

seed = 1234;

randn('state', seed );
rand('twister', seed+1 );


%you will NEVER need more than a few hundred epochs unless you are doing
%something very wrong.  Here 'epoch' means parameter update, not 'pass over
%the training set'.
maxepoch = 500;


%CURVES
%%%%%%%%%%%%%%%%%
%this dataset (by Ruslan Salakhutdinov) is available here: http://www.cs.toronto.edu/~jmartens/digs3pts_1.mat

tmp = load('digs3pts_1.mat');
indata = tmp.bdata';
%outdata = tmp.bdata;
intest = tmp.bdatatest';
%outtest = tmp.bdatatest;
clear tmp

perm = randperm(size(indata,2));
indata = indata( :, perm );

%it's an auto-encoder so output is input
outdata = indata;
outtest = intest;


runName = 'HFtestrun2';

runDesc = ['seed = ' num2str(seed) ', enter anything else you want to remember here' ];

%next try using autodamp = 0 for rho computation.  both for version 6 and
%versions with rho and cg-backtrack computed on the training set


layersizes = [400 200 100 50 25 6 25 50 100 200 400];
%Note that the code layer uses linear units
layertypes = {'logistic', 'logistic', 'logistic', 'logistic', 'logistic', 'linear', 'logistic', 'logistic', 'logistic', 'logistic', 'logistic', 'logistic'};

resumeFile = [];

paramsp = [];
Win = [];
bin = [];
%[Win, bin] = loadPretrainedNet_curves;

numchunks = 4
numchunks_test = 4;

mattype = 'gn'; %Gauss-Newton.  The other choices probably won't work for whatever you're doing
%mattype = 'hess';
%mattype = 'empfish';

rms = 0;

hybridmode = 1;

%decay = 1.0;
decay = 0.95;

%jacket = 0;
%this enables Jacket mode for the GPU
jacket = 1;

errtype = 'L2'; %report the L2-norm error (in addition to the quantity actually being optimized, i.e. the log-likelihood)

%standard L_2 weight-decay:
weightcost = 2e-5
%weightcost = 0


nnet_train_2( runName, runDesc, paramsp, Win, bin, resumeFile, maxepoch, indata, outdata, numchunks, intest, outtest, numchunks_test, layersizes, layertypes, mattype, rms, errtype, hybridmode, weightcost, decay, jacket);