void LayerNet::an1 ( TrainingSet *tptr , struct LearnParams *lptr )
{
int itry, user_quit ;
long seed ;
double best_err ;
char msg[80] ;
LayerNet *worknet, *bestnet ;
/*
Allocate scratch memory
*/
MEMTEXT ( "AN1::learn new worknet, bestnet" ) ;
worknet = new LayerNet ( model , outmod , outlin , nin , nhid1 , nhid2 ,
nout , 0 , 0 ) ;
bestnet = new LayerNet ( model , outmod , outlin , nin , nhid1 , nhid2 ,
nout , 0 , 1 ) ;
if ((worknet == NULL) || (! worknet->ok)
|| (bestnet == NULL) || (! bestnet->ok)) {
memory_message ( "to learn" ) ;
if (worknet != NULL)
delete worknet ;
if (bestnet != NULL)
delete bestnet ;
errtype = 0 ;
return ;
}
best_err = 1.e30 ;
for (itry=1 ; itry<=lptr->retries+1 ; itry++) {
user_quit = anneal1 ( tptr , lptr , worknet , 1 , itry ) ;
if (neterr < best_err) {
best_err = neterr ;
copy_weights ( bestnet , this ) ;
}
sprintf ( msg , "Try %d err=%lf best=%lf", itry, neterr, best_err ) ;
normal_message ( msg ) ;
if (user_quit || (neterr < lptr->quit_err))
break ;
seed = flrand() - (long) (itry * 97) ; // Insure new seed for anneal
sflrand ( seed ) ;
zero_weights () ; // Retry random
}
copy_weights ( this , bestnet ) ;
neterr = best_err ;
MEMTEXT ( "AN1::learn delete worknet, bestnet" ) ;
delete worknet ;
delete bestnet ;
return ;
}
void Network::show_confusion ()
{
int i ;
char *msg ;
MEMTEXT ( "CONFUSE:show msg" ) ;
if ((msg = (char *) MALLOC ( (nout+1) * 5 + 11 )) == NULL ) {
memory_message ( "to SHOW CONFUSION" ) ;
return ;
}
strcpy ( msg , "Confusion:" ) ;
for (i=0 ; i<nout ; i++)
sprintf ( msg+5*i+10 , "%5d" , confusion[i] ) ;
sprintf ( msg+5*nout+10, "%5d", confusion[nout] ) ;
msg[5*nout+15] = 0 ;
normal_message ( msg ) ;
FREE ( msg ) ;
return ;
}
void LayerNet::anx_dd ( TrainingSet *tptr , struct LearnParams *lptr )
{
int itry, n_escape, n_retry, bad_count, new_record, refined ;
long seed ;
double err, prev_err, best_err, start_of_loop_error, best_inner_error ;
double initial_accuracy, final_accuracy ;
char msg[80] ;
LayerNet *worknet, *worknet2, *bestnet ;
n_escape = n_retry = 0 ;
/*
Allocate scratch memory
*/
MEMTEXT ( "ANX_DD::learn new worknet, bestnet" ) ;
worknet = new LayerNet ( model , outmod , outlin , nin , nhid1 , nhid2 ,
nout , 0 , 0 ) ;
bestnet = new LayerNet ( model , outmod , outlin , nin , nhid1 , nhid2 ,
nout , 0 , 1 ) ;
if ((worknet == NULL) || (! worknet->ok)
|| (bestnet == NULL) || (! bestnet->ok)) {
memory_message ( "to learn" ) ;
if (worknet != NULL)
delete worknet ;
if (bestnet != NULL)
delete bestnet ;
errtype = 0 ;
return ;
}
if ((lptr->method == METHOD_AN2_CJ) || (lptr->method == METHOD_AN2_LM)) {
worknet2 = new LayerNet ( model , outmod , outlin , nin , nhid1 , nhid2 ,
nout , 0 , 0 ) ;
if ((worknet2 == NULL) || (! worknet2->ok)) {
if (worknet2 != NULL)
delete worknet2 ;
delete worknet ;
delete bestnet ;
memory_message ( "to learn" ) ;
errtype = 0 ;
return ;
}
}
else
worknet2 = NULL ;
/*
Start by annealing around the starting weights. These will be zero if the
net was just created. If it was restored or partially trained already,
they will be meaningful. Anneal1 guarantees that it will not return all
zero weights if there is at least one hidden layer, even if that means
that the error exceeds the amount that could be attained by all zeros.
*/
best_err = best_inner_error = 1.e30 ;
if ((lptr->method == METHOD_AN1_CJ) || (lptr->method == METHOD_AN1_LM))
anneal1 ( tptr , lptr , worknet , 1 , 0 ) ;
else if ((lptr->method == METHOD_AN2_CJ) || (lptr->method == METHOD_AN2_LM))
anneal2 ( tptr , lptr , worknet , worknet2 , 1 ) ;
/*
Do direct descent optimization, finding local minimum.
Then anneal to break out of it. If successful, loop back up to
do direct descent again. Otherwise restart totally random.
*/
bad_count = 0 ; // Handles flat local mins
refined = 0 ; // Did we ever refine to high resolution? Not yet.
new_record = 0 ; // Refine every time a new inner error record set
initial_accuracy = pow ( 10.0 , -lptr->cj_acc ) ;
final_accuracy = initial_accuracy * pow ( 10.0 , -lptr->cj_refine ) ;
for (itry=1 ; ; itry++) {
if (neterr < best_err) { // Keep track of best
copy_weights ( bestnet , this ) ;
best_err = neterr ;
}
sprintf ( msg , "Try %d (best=%lf):", itry, best_err ) ;
normal_message ( msg ) ;
if (neterr <= lptr->quit_err)
break ;
start_of_loop_error = neterr ;
if ((lptr->method == METHOD_AN1_CJ) || (lptr->method == METHOD_AN2_CJ))
err = conjgrad ( tptr , 32767 , initial_accuracy ,
lptr->quit_err , lptr->cj_progress ) ;
else if ((lptr->method==METHOD_AN1_LM) || (lptr->method==METHOD_AN2_LM))
err = lev_marq ( tptr , 32767 , initial_accuracy ,
lptr->quit_err , lptr->cj_progress ) ;
neterr = fabs ( err ) ; // err<0 if user pressed ESCape
sprintf ( msg , " Gradient err=%lf", neterr ) ;
progress_message ( msg ) ;
if (neterr < best_err) { // Keep track of best
copy_weights ( bestnet , this ) ;
best_err = neterr ;
}
if (err <= lptr->quit_err) { // err<0 if user pressed ESCape
if (err < -1.e29) // or insufficient memory
printf ( "\nInsufficient memory for gradient learning." ) ;
break ;
}
seed = flrand() - (long) (itry * 97) ; // Insure new seed for anneal
sflrand ( seed ) ;
prev_err = neterr ; // So we can see if anneal helped
if ((lptr->method == METHOD_AN1_CJ) || (lptr->method == METHOD_AN1_LM))
anneal1 ( tptr , lptr , worknet , 0 , itry ) ;
else if ((lptr->method==METHOD_AN2_CJ) || (lptr->method==METHOD_AN2_LM))
anneal2 ( tptr , lptr , worknet , worknet2 , 0 ) ;
sprintf ( msg , " Anneal err=%lf", neterr ) ;
progress_message ( msg ) ;
if (neterr < best_err) { // Keep track of best
copy_weights ( bestnet , this ) ;
best_err = neterr ;
}
if (best_err <= lptr->quit_err)
break ;
if (neterr < best_inner_error) { // Keep track of best inner for refine
best_inner_error = neterr ;
new_record = 1 ; // Tells us to refine
}
if ((prev_err - neterr) > 1.e-7) { // Did we break out of local min?
if ((start_of_loop_error - neterr) < 1.e-3)
++bad_count ; // Avoid many unprofitable iters
else
bad_count = 0 ;
if (bad_count < 4) {
++n_escape ; // Count escapes from local min
continue ; // Escaped, so gradient learn again
}
}
/*
After first few tries, and after each inprovement thereafter, refine
to high resolution
*/
if ((itry-n_escape >= lptr->cj_pretries) && (new_record || ! refined)) {
if (! refined) { // If refining the best of the pretries
copy_weights ( this , bestnet ) ; // Get that net
neterr = best_err ;
}
refined = 1 ; // Only force refine once
new_record = 0 ; // Reset new inner error record flag
progress_message ( " REFINING" ) ;
if ((lptr->method == METHOD_AN1_CJ) || (lptr->method == METHOD_AN2_CJ))
err = conjgrad ( tptr , 0 , final_accuracy ,
lptr->quit_err , lptr->cj_progress ) ;
else if ((lptr->method==METHOD_AN1_LM)|| (lptr->method==METHOD_AN2_LM))
err = lev_marq ( tptr , 0 , final_accuracy ,
lptr->quit_err , lptr->cj_progress ) ;
neterr = fabs ( err ) ; // err<0 if user pressed ESCape
sprintf ( msg , " Attained err=%lf", neterr ) ;
progress_message ( msg ) ;
if (neterr < best_err) { // Keep track of best
copy_weights ( bestnet , this ) ;
best_err = neterr ;
}
}
if (++n_retry > lptr->retries)
break ;
progress_message ( " RESTART" ) ;
zero_weights () ; // Failed to break out, so retry random
seed = flrand() - (long) (itry * 773) ; // Insure new seed for anneal
sflrand ( seed ) ;
if ((lptr->method == METHOD_AN1_CJ) || (lptr->method == METHOD_AN1_LM))
anneal1 ( tptr , lptr , worknet , 1 , itry ) ;
else if ((lptr->method==METHOD_AN2_CJ) || (lptr->method==METHOD_AN2_LM))
anneal2 ( tptr , lptr , worknet , worknet2 , 1 ) ;
}
FINISH:
copy_weights ( this , bestnet ) ;
neterr = best_err ;
MEMTEXT ( "AN1_DD::learn delete worknet, bestnet" ) ;
delete worknet ;
delete bestnet ;
sprintf ( msg , "%d successful escapes, %d retries", n_escape, n_retry ) ;
normal_message ( msg ) ;
return ;
}
//*******************************************************************
// WinMain - Neural main
//
// parameters:
// hInstance - The instance of this instance of this
// application.
// hPrevInstance - The instance of the previous instance
// of this application. This will be 0
// if this is the first instance.
// lpszCmdLine - A long pointer to the command line that
// started this application.
// cmdShow - Indicates how the window is to be shown
// initially. ie. SW_SHOWNORMAL, SW_HIDE,
// SW_MIMIMIZE.
//
// returns:
// wParam from last message.
//
//*******************************************************************
int PASCAL WinMain(HINSTANCE hInstance, HINSTANCE hPrevInstance,
LPSTR lpszCmdLine, int cmdShow)
{
/*
Declarations of local variables
*/
int control_file_number = -1 ; // Stack pointer for control files
FILE *control_files[MAX_CONTROL_FILES] ; // This is the stack
char *control_line ; // User's commands here
char *command, *rest ; // Pointers to its command and parameter parts
int n_command, n_rest ; // Lengths of those parts
int net_model = -1 ; // Network model (see NETMOD_? in CONST.H)
int out_model = -1 ; // Output model (see OUTMOD_? in CONST.H)
int n_inputs = -1 ; // Number of input neurons
int n_outputs = -1 ; // Number of output neurons
int n_hidden1 = -1 ; // Number of hidden layer one neurons
int n_hidden2 = -1 ; // Ditto layer 2 (0 if just one hidden layer)
TrainingSet *tset = NULL ; // Training set here
Network *network = NULL ; // Network here
struct LearnParams learn_params ; // General learning parameters
struct AnnealParams anneal_params ; // Simulated annealing parameters
struct GenInitParams geninit_params ; // Genetic initialization parameters
struct KohParams koh_params ; // Kohonen parameters
int classif_output = -1 ; // Current class (0=reject) for classif training
char out_file[80] = "" ; // File for EXECUTE output
float threshold ; // CLASSIFY confusion reject cutoff
char resp_file[80]=""; // file for initializing output neuron's name
char train_file[80]="";
/*
Miscellaneous variables
*/
int i, n, m ;
float p ;
char *msg ;
FILE *fp ;
unsigned long me,mc;
char *fname;
char *control;
#if VERSION_16_BIT
if (sizeof(int) > 2) {
printf ( "\nRecompile with VERSION_16_BIT set to 0 in CONST.H" ) ;
exit ( 1 ) ;
}
#else
if (sizeof(int) < 4) {
printf ( "\nRecompile with VERSION_16_BIT set to 1 in CONST.H" ) ;
exit ( 1 ) ;
}
#endif
printf ( "\nNEURAL SYSTEM - Program to train and test neural networks" ) ;
if (argc>1)
{
strcpy(fname,argv[1]);
}
/*
Process command line parameters
*/
mem_name[0] = 0 ; // Default is no memory allocation file
/*
if (strlen ( mem_name )) {
strcat ( mem_name , ":mem.log" ) ;
fp = fopen ( mem_name , "wt" ) ;
if (fp == NULL) {
printf ( "\nCannot open debugging file %s", mem_name ) ;
exit ( 1 ) ;
}
fclose ( fp ) ;
mem_log = 1 ;
}
else
mem_log = 0 ;
*/
mem_log = 0 ;
mem_used = 0 ;
/*
Initialize defaults
*/
learn_params.init = -1 ;
learn_params.quit_err = 0.0 ;
learn_params.retries = 32767 ;
anneal_params.temps0 = 3 ;
anneal_params.temps = 4 ;
anneal_params.iters0 = 50 ;
anneal_params.iters = 20 ;
anneal_params.setback0 = 50 ;
anneal_params.setback = 20 ;
anneal_params.start0 = 3.0 ;
anneal_params.start = 4.0 ;
anneal_params.stop0 = 1.0 ;
anneal_params.stop = 0.02 ;
geninit_params.pool = 50 ;
geninit_params.gens = 3 ;
geninit_params.climb = 0 ;
geninit_params.overinit = 1.5 ;
geninit_params.pcross = 0.8 ;
geninit_params.pmutate = 0.0001 ;
koh_params.normalization = 0 ; // 0=multiplicative, 1=Z
koh_params.learn_method = 1 ; // 0=additive, 1=subtractive
koh_params.rate = 0.4 ; // learning rate
koh_params.reduction = 0.99 ; // learning rate reduction
learn_params.ap = &anneal_params ;
learn_params.gp = &geninit_params ;
learn_params.kp = &koh_params ;
act_func_init () ; // Initialize interpolation table for activation function
MEMTEXT ( "NEURAL: control_line, msg" ) ;
if (((control_line = (char *) MALLOC ( CONTROL_LINE_LENGTH+1 )) == NULL)
|| ((msg = (char *) MALLOC ( CONTROL_LINE_LENGTH+1 )) == NULL)) {
printf ( "\nInsufficient memory" ) ;
exit ( 1 ) ;
}
/*
Main loop processes all commands
*/
for (;;) {
if (argv[1])
{
strcpy(control_line,"CONTROL:");
strcat(control_line,fname);
//printf("%s\n",control_line);
argv[1]=NULL;
}
else
get_control_line ( control_line , &control_file_number, control_files ) ;
split_control_line ( control_line , &command , &n_command ,
&rest , &n_rest ) ;
if (! n_command) {
if (n_rest) {
sprintf ( msg , "No colon after command: %s", rest ) ;
error_message ( msg ) ;
}
continue ;
}
sprintf ( msg , "%s : %s", command, rest ) ;
normal_message ( msg ) ;
/*
Act on the command
*/
if (! strcmp ( command , "QUIT" ))
break ;
if (! strcmp ( command , "CONTROL" )) {
stack_control_file (rest, &control_file_number, control_files) ;
continue ;
}
if (! strcmp ( command , "NETWORK MODEL" )) {
// Multi layer network
if (! strcmp ( rest , "LAYER" ))
n = NETMOD_LAYER ;
// Kohonen network
else if (! strcmp ( rest , "KOHONEN" ))
n = NETMOD_KOH ;
// Hopfield network
else if (! strcmp ( rest , "HOPFIELD" ))
n = NETMOD_HOP ;
// Bidirectionnal associative memory network
else if (! strcmp ( rest , "BAM" ))
n = NETMOD_BAM ;
else {
sprintf ( msg , "Illegal NETWORK MODEL: %s", rest ) ;
error_message ( msg ) ;
continue ;
}
if (net_model == n)
continue ;
if (ok_to_clear_weights( &network )) {
net_model = n ;
learn_params.init = -1 ;
}
else
warning_message ( "Command aborted" ) ;
continue ;
}
if (! strcmp ( command , "OUTPUT MODEL" )) {
if (! strcmp ( rest , "CLASSIFY" ))
n = OUTMOD_CLASSIFY ;
else if (! strcmp ( rest , "AUTO" ))
n = OUTMOD_AUTO ;
else if (! strcmp ( rest , "GENERAL" ))
n = OUTMOD_GENERAL ;
else {
sprintf ( msg , "Illegal OUTPUT MODEL: %s", rest ) ;
error_message ( msg ) ;
continue ;
}
if (out_model == n)
continue ;
if ((ok_to_clear_tset( &tset )) && (ok_to_clear_weights( &network)))
out_model = n ;
else
warning_message ( "Command aborted" ) ;
continue ;
}
if (! strcmp ( command , "N INPUTS" )) {
m = sscanf ( rest , "%d" , &n ) ;
if ((m <= 0) || (n <= 0) || (n > MAX_INPUTS)) {
sprintf ( msg , "Illegal N INPUTS: %s", rest ) ;
error_message ( msg ) ;
continue ;
}
if (n_inputs == n)
continue ;
if ((ok_to_clear_tset( &tset)) && (ok_to_clear_weights(&network)))
n_inputs = n ;
else
warning_message ( "Command aborted" ) ;
continue ;
}
if (! strcmp ( command , "N OUTPUTS" )) {
m = sscanf ( rest , "%d" , &n ) ;
if ((m <= 0) || (n <= 0) || (n > MAX_OUTPUTS)) {
sprintf ( msg , "Illegal N OUTPUTS: %s", rest ) ;
error_message ( msg ) ;
continue ;
}
if (n_outputs == n)
continue ;
if ((ok_to_clear_tset( &tset)) && (ok_to_clear_weights(&network)))
n_outputs = n ;
else
warning_message ( "Command aborted" ) ;
continue ;
}
if (! strcmp ( command , "N HIDDEN1" )) {
m = sscanf ( rest , "%d" , &n ) ;
if ((m <= 0) || (n < 0) || (n > MAX_HIDDEN)) {
sprintf ( msg , "Illegal N HIDDEN1: %s", rest ) ;
error_message ( msg ) ;
continue ;
}
if (n_hidden1 == n)
continue ;
if (ok_to_clear_weights( &network ))
n_hidden1 = n ;
else
warning_message ( "Command aborted" ) ;
continue ;
}
if (! strcmp ( command , "N HIDDEN2" )) {
m = sscanf ( rest , "%d" , &n ) ;
if ((m <= 0) || (n < 0) || (n > MAX_HIDDEN)) {
sprintf ( msg , "Illegal N HIDDEN2: %s", rest ) ;
error_message ( msg ) ;
continue ;
}
if (n && ! n_hidden1) {
error_message ( "N HIDDEN2 must be 0 if N HIDDEN1 IS 0." ) ;
continue ;
}
if (n_hidden2 == n)
continue ;
if (ok_to_clear_weights( &network ))
n_hidden2 = n ;
else
warning_message ( "Command aborted" ) ;
continue ;
}
if (! strcmp ( command , "TRAIN" )) {
if ((out_model == OUTMOD_AUTO) && (n_outputs != n_inputs)) {
warning_message ( "Setting N OUTPUTS = N INPUTS" ) ;
n_outputs = n_inputs ;
}
if (out_model <= 0)
error_message ( "TRAIN used before OUTPUT MODEL set." ) ;
else if (n_inputs <= 0)
error_message ( "TRAIN used before N INPUTS set." ) ;
else if (n_outputs <= 0)
error_message ( "TRAIN used before N OUTPUTS set." ) ;
else if ((net_model == NETMOD_HOP) && (n_inputs != n_outputs))
error_message("HOPFIELD netowork requires INPUTS = OUTPUTS.");
else if ((net_model == NETMOD_BAM) && (out_model != OUTMOD_GENERAL))
error_message("BAM network requires AUTO output mode.");
else if ((net_model == NETMOD_HOP) && (out_model != OUTMOD_AUTO))
error_message("HOFIELD network requires AUTO output mode.");
else if ((net_model != NETMOD_KOH) && (out_model == OUTMOD_CLASSIFY)
&& (classif_output < 0))
error_message( "CLASSIFY output mode but CLASSIFY OUTPUT not set.");
else if ((net_model == NETMOD_KOH) && (out_model != OUTMOD_CLASSIFY))
error_message( "KOHONEN network requires CLASSIFY output mode.");
else {
if (tset == NULL) {
MEMTEXT ( "NEURAL: new tset" ) ;
tset = new TrainingSet ( out_model , n_inputs , n_outputs ) ;
}
tset->train ( rest , classif_output ) ;
strcpy(train_file,rest);
}
continue ;
}
if (check_anneal ( command , rest , &anneal_params ))
continue ;
if (check_genetic ( command , rest , &geninit_params ))
continue ;
if (check_kohonen ( command , rest , &koh_params , &network ))
continue ;
if (check_learn_params ( command , rest , &learn_params , net_model ))
continue ;
if (! strcmp ( command , "LEARN" )) {
if ((tset == NULL) || (tset->ntrain == 0)) {
error_message ( "Cannot LEARN; No training set exists." ) ;
continue ;
}
if ((net_model == NETMOD_KOH) && (out_model != OUTMOD_CLASSIFY)) {
error_message( "KOHONEN network requires CLASSIFY output mode.");
continue ;
}
if (learn_params.init < 0) {
error_message( "Initialization method not set.");
continue ;
}
if (network == NULL)
{
if (net_model == NETMOD_LAYER)
{
if (n_hidden1 < 0)
{
error_message ( "LEARN used before N HIDDEN1 set." ) ;
continue ;
}
else if (n_hidden2 < 0)
{
error_message ( "LEARN used before N HIDDEN2 set." ) ;
continue ;
}
else
{
MEMTEXT ( "NEURAL: new LayerNet" ) ;
network = new LayerNet ( out_model , n_inputs , n_hidden1 ,
n_hidden2 , n_outputs , 1 , 1 ) ;
}
}
else if (net_model == NETMOD_KOH)
{
MEMTEXT ( "NEURAL: new KohNet" ) ;
network = new KohNet ( n_inputs , n_outputs ,
&koh_params , 1 , 1 ) ;
}
else if (net_model == NETMOD_HOP)
{
MEMTEXT ( "NEURAL: new HopNet" );
network = new HopNet (n_inputs,n_outputs, 1,1);
}
else if (net_model == NETMOD_BAM)
{
MEMTEXT ("NEURAL: new BamNet");
network = new LayerNet ( out_model , n_inputs , n_hidden1 ,
n_hidden2 , n_outputs , 1 , 1 ) ;
}
}
if ((network == NULL) || (! network->ok)) { // Malloc failure?
memory_message ( "to create network." ) ;
if (network != NULL) {
delete network ;
network = NULL ;
}
continue ;
}
normal_message("Learning...\n");
network->learn ( tset , &learn_params ) ;
normal_message("End of Learning\n");
if (network->neterr > 0.999999) { // Indicates massive failure
MEMTEXT ( "NEURAL: learn failure delete network" ) ;
delete network ;
network = NULL ;
}
else {
sprintf ( msg , "Final error = %.4lf%% of max possible",
100.0 * network->neterr ) ;
normal_message ( msg ) ;
}
continue ;
}
if (! strcmp ( command , "SAVE WEIGHTS" )) {
if (network == NULL)
error_message ( "There are no learned weights to save." ) ;
else
wt_save ( network , net_model , 0 , rest ) ;
continue ;
}
if (! strcmp ( command , "RESTORE WEIGHTS" )) {
if (network != NULL) {
MEMTEXT ( "NEURAL: delete network for restore" ) ;
delete network ;
network = NULL ;
}
network = wt_restore ( rest , &net_model ) ;
if (network == NULL)
continue ;
if (tset != NULL) {
if ((tset->nin != network->nin)
|| (tset->nout != network->nout)
|| (tset->outmod != network->outmod)) {
error_message ( "Network conflicts with existing training set.");
continue ;
}
}
out_model = network->outmod ;
n_inputs = network->nin ;
n_outputs = network->nout ;
if (net_model == NETMOD_LAYER) {
n_hidden1 = ((LayerNet*) network)->nhid1 ;
n_hidden2 = ((LayerNet*) network)->nhid2 ;
}
if (net_model == NETMOD_KOH)
koh_params.normalization = ((KohNet *) network)->normalization ;
learn_params.init = -1 ;
continue ;
}
if (! strcmp ( command , "CLEAR TRAINING" )) {
if (tset != NULL) {
MEMTEXT ( "NEURAL: delete tset" ) ;
delete tset ;
tset = NULL ;
}
continue ;
}
if (! strcmp ( command , "CLEAR WEIGHTS" )) {
if (network != NULL) {
MEMTEXT ( "NEURAL: delete network" ) ;
delete network ;
network = NULL ;
}
continue ;
}
if (! strcmp ( command , "CLASSIFY OUTPUT" )) {
if (net_model == NETMOD_KOH) {
error_message ( "Cannot specify output for KOHONEN model." ) ;
continue ;
}
if (n_outputs < 0) {
error_message ( "CLASSIFY OUTPUT used before N OUTPUTS set." ) ;
continue ;
}
if (out_model != OUTMOD_CLASSIFY) {
error_message
( "CLASSIFY OUTPUT only valid when OUTPUT MODEL:CLASSIFY" ) ;
continue ;
}
m = sscanf ( rest , "%d" , &n ) ;
if ((m <= 0) || (n < 0)) {
sprintf ( msg , "Illegal CLASSIFY OUTPUT: %s", rest ) ;
error_message ( msg ) ;
}
else if (n > n_outputs) {
sprintf ( msg , "CLASSIFY OUTPUT (%d) exceeds N OUTPUTS (%d)",
n, n_outputs ) ;
error_message ( msg ) ;
}
else
classif_output = n ;
continue ;
}
if (! strcmp ( command , "OUTPUT FILE" )) {
strcpy ( out_file , rest ) ;
continue ;
}
if (! strcmp ( command , "EXECUTE" ))
{
if (network == NULL)
error_message ( "There is no trained network" ) ;
else
{
network->execute_from_file ( rest , out_file) ;
continue ;
}
}
if (! strcmp ( command , "TEST NETWORK" ))
{
if (network == NULL)
error_message ( "There is no trained network" ) ;
else
{
network->test_from_file ( rest ,out_file,net_model) ;
continue ;
}
}
if (! strcmp ( command , "CLASSIFY" )) {
if (network == NULL)
error_message ( "There is no trained network" ) ;
else if (out_model != OUTMOD_CLASSIFY)
error_message ( "CLASSIFY valid only in CLASSIFY output mode" ) ;
else
network->classify_from_file ( rest , threshold ) ;
continue ;
}
if (! strcmp ( command , "RESET CONFUSION" )) {
if (network == NULL)
error_message ( "There is no trained network" ) ;
else
network->reset_confusion () ;
continue ;
}
if (! strcmp ( command , "CONFUSION THRESHOLD" )) {
p = atof ( rest ) ;
if ((p < 0.0) || (p > 100.0)) {
sprintf ( msg , "Illegal CONFUSION THRESHOLD: %s", rest ) ;
error_message ( msg ) ;
}
else
threshold = p / 100.0 ;
continue ;
}
if (! strcmp ( command , "SHOW CONFUSION" )) {
if (network == NULL)
error_message ( "There is no trained network" ) ;
else if (out_model != OUTMOD_CLASSIFY)
error_message ( "CONFUSION valid only in CLASSIFY output mode" ) ;
else
network->show_confusion () ;
continue ;
}
if (! strcmp ( command , "SAVE CONFUSION" )) {
if (network == NULL)
error_message ( "There is no trained network" ) ;
else if (out_model != OUTMOD_CLASSIFY)
error_message ( "CONFUSION valid only in CLASSIFY output mode" ) ;
else
network->save_confusion ( rest ) ;
continue ;
}
sprintf ( msg , "Unknown command: %s", command ) ;
error_message ( msg ) ;
} // Endless command loop
MEMTEXT ( "NEURAL: control_line, msg" ) ;
FREE ( control_line ) ;
FREE ( msg ) ;
MEMCLOSE () ;
exit ( 0 ) ;
}
void LayerNet::ssg (
TrainingSet *tptr , // Training set to use
struct LearnParams *lptr , // User's general learning parameters
int use_grad // SS if zero, else SSG
)
{
int itry, user_quit, n, n_grad ;
long seed ;
double best_err, *work1, *work2, *grad, *avg_grad ;
char msg[80] ;
LayerNet *worknet1, *worknet2, *bestnet ;
/*
Allocate network scratch memory
*/
MEMTEXT ( "SSG::new 2 worknets, bestnet" ) ;
worknet1 = new LayerNet ( model , outmod , outlin , nin , nhid1 , nhid2 ,
nout , 0 , 0 ) ;
worknet2 = new LayerNet ( model , outmod , outlin , nin , nhid1 , nhid2 ,
nout , 0 , 0 ) ;
bestnet = new LayerNet ( model , outmod , outlin , nin , nhid1 , nhid2 ,
nout , 0 , 1 ) ;
if ((worknet1 == NULL) || (! worknet1->ok)
|| (worknet2 == NULL) || (! worknet2->ok)
|| (bestnet == NULL) || (! bestnet->ok)) {
memory_message ( "to learn" ) ;
if (worknet1 != NULL)
delete worknet1 ;
if (worknet2 != NULL)
delete worknet2 ;
if (bestnet != NULL)
delete bestnet ;
errtype = 0 ;
return ;
}
/*
Allocate gradient work memory.
Work1 is used for hidden layer 2 deltas in REAL model, and output
activation partial derivatives and deltas in all COMPLEX models.
Work2 is output deltas in REAL model, error difference in COMPLEX models.
*/
if (use_grad) {
if (nhid2) // Must be REAL model if this is true
n = nhid2 ;
else if (model == NETMOD_COMPLEX_INPUT)
n = nhid1 ? nout * 2 + nhid1 * 2 : nout * 2 ;
else if (model == NETMOD_COMPLEX_HIDDEN)
n = nout * 4 + nhid1 * 4 ;
else if (model == NETMOD_COMPLEX)
n = nhid1 ? nout * 6 + nhid1 * 4 : nout * 4 ;
else
n = 0 ;
if (n) {
MEMTEXT ( "SSG::work1" ) ;
work1 = (double *) MALLOC ( n * sizeof(double) ) ;
if (work1 == NULL) {
memory_message ( "to learn" ) ;
delete worknet1 ;
delete worknet2 ;
delete bestnet ;
errtype = 0 ;
return ;
}
}
else
work1 = NULL ;
if (nhid1 == 0) // No hidden layer
n_grad = nout * nin_n ;
else if (nhid2 == 0) // One hidden layer
n_grad = nhid1 * nin_n + nout * nhid1_n ;
else // Two hidden layers
n_grad = nhid1 * nin_n + nhid2 * nhid1_n + nout * nhid2_n ;
MEMTEXT ( "SSG::3 work vectors" ) ;
work2 = (double *) MALLOC ( nout_n * sizeof(double) ) ;
grad = (double *) MALLOC ( n_grad * sizeof(double) ) ;
avg_grad = (double *) MALLOC ( n_grad * sizeof(double) ) ;
if ((work2 == NULL) || (grad == NULL) || (avg_grad == NULL)) {
if (work1 != NULL)
FREE ( work1 ) ;
if (work2 != NULL)
FREE ( work2 ) ;
if (grad != NULL)
FREE ( grad ) ;
if (avg_grad != NULL)
FREE ( avg_grad ) ;
memory_message ( "to learn" ) ;
delete worknet1 ;
delete worknet2 ;
delete bestnet ;
errtype = 0 ;
return ;
}
}
else
work1 = work2 = grad = avg_grad = NULL ;
best_err = 1.e30 ;
for (itry=1 ; itry<=lptr->retries+1 ; itry++) {
user_quit = ssg_core ( tptr , lptr , worknet1 , worknet2 ,
work1 , work2 , grad , avg_grad , n_grad ) ;
if (neterr < best_err) {
best_err = neterr ;
copy_weights ( bestnet , this ) ;
}
sprintf ( msg , "Try %d err=%lf best=%lf", itry, neterr, best_err ) ;
normal_message ( msg ) ;
if (user_quit || (neterr < lptr->quit_err))
break ;
seed = flrand() - (long) (itry * 97) ; // Insure new seed for anneal
sflrand ( seed ) ;
zero_weights () ; // Retry random
}
copy_weights ( this , bestnet ) ;
neterr = best_err ;
MEMTEXT ( "AN1::learn delete 2 worknets, bestnet" ) ;
delete worknet1 ;
delete worknet2 ;
delete bestnet ;
if (use_grad) {
if (work1 != NULL) {
MEMTEXT ( "SSG::work1" ) ;
FREE ( work1 ) ;
}
MEMTEXT ( "SSG::3 work vectors" ) ;
FREE ( work2 ) ;
FREE ( grad ) ;
FREE ( avg_grad) ;
}
return ;
}
int LayerNet::ssg_core (
TrainingSet *tptr , // Training set to use
struct LearnParams *lptr , // User's general learning parameters
LayerNet *avgnet , // Work area used to keep average weights
LayerNet *bestnet , // And the best so far
double *work1 , // Gradient work vector
double *work2 , // Ditto
double *grad , // Ditto
double *avg_grad , // Ditto
int n_grad // Length of above vectors
)
{
int ntemps, niters, setback, reg, nvars, user_quit ;
int i, iter, itemp, n_good, n_bad, use_grad ;
char msg[80] ;
double tempmult, temp, fval, bestfval, starttemp, stoptemp, fquit ;
double avg_func, new_fac, gradlen, grad_weight, weight_used ;
enum RandomDensity density ;
SingularValueDecomp *sptr ;
struct AnnealParams *aptr ; // User's annealing parameters
aptr = lptr->ap ;
ntemps = aptr->temps0 ;
niters = aptr->iters0 ;
setback = aptr->setback0 ;
starttemp = aptr->start0 ;
stoptemp = aptr->stop0 ;
if (aptr->random0 == ANNEAL_GAUSSIAN)
density = NormalDensity ;
else if (aptr->random0 == ANNEAL_CAUCHY)
density = CauchyDensity ;
if (! (ntemps * niters))
return 0 ;
/*
Initialize other local parameters. Note that there is no sense using
regression if there are no hidden layers.
*/
use_grad = (grad != NULL) ;
fquit = lptr->quit_err ;
reg = nhid1 ;
/*
Allocate the singular value decomposition object for REGRESS.
Also allocate a work area for REGRESS to preserve matrix.
*/
if (reg) { // False if no hidden layers
if (nhid2 == 0) // One hidden layer
nvars = nhid1_n ;
else // Two hidden layers
nvars = nhid2_n ;
i = (model == NETMOD_COMPLEX) ? 2 * tptr->ntrain : tptr->ntrain ;
if (i < nvars) {
warning_message ( "Too few training sets for regression." ) ;
reg = 0 ;
}
else {
MEMTEXT ( "SSG: new SingularValueDecomp" ) ;
sptr = new SingularValueDecomp ( i , nvars , 1 ) ;
if ((sptr == NULL) || ! sptr->ok) {
memory_message (
"for SS(G) with regression. Using total randomization.");
if (sptr != NULL)
delete sptr ;
reg = 0 ;
}
}
}
/*
For the basic algorithm, we will keep the current 'average' network
weight set in avgnet. This will be the moving center about which the
perturbation is done.
Although not directly related to the algorithm itself, we keep track
of the best network ever found in bestnet. That is what the user
will get at the end.
*/
copy_weights ( bestnet , this ) ; // Current weights are best so far
copy_weights ( avgnet , this ) ; // Center of perturbation
bestfval = trial_error ( tptr ) ;
/*
If this is being used to initialize the weights, make sure that they are
not identically zero. Do this by setting bestfval huge so that
SOMETHING is accepted later.
*/
if (nhid1) {
i = nhid1 * nin_n ;
while (i--) {
if (fabs(hid1_coefs[i]) > 1.e-10)
break ;
}
if (i < 0)
bestfval = 1.e30 ;
}
/*
Initialize by cumulating a bunch of points
*/
normal_message ( "Initializing..." ) ;
avg_func = 0.0 ; // Mean function around center
if (use_grad) {
for (i=0 ; i<n_grad ; i++) // Zero the mean gradient
avg_grad[i] = 0.0 ;
}
for (iter=0 ; iter<niters ; iter++) { // Initializing iterations
perturb ( avgnet , this , starttemp , reg , density ) ; // Move point
if (reg) // If using regression, estimate
fval = regress ( tptr , sptr ) ; // out weights now, ignore fval
if (use_grad) // Also need gradient?
fval = gradient ( tptr , work1 , work2 , grad ) ; // fval redundant
else if (! reg) // If reg we got fval from regress
fval = trial_error ( tptr ) ;
avg_func += fval ; // Cumulate mean function
if (use_grad) { // Also need gradient?
for (i=0 ; i<n_grad ; i++) // Cumulate mean gradient
avg_grad[i] += grad[i] ;
}
if (fval < bestfval) { // If this iteration improved
bestfval = fval ; // then update the best so far
copy_weights ( bestnet , this ) ; // Keep the network
if (bestfval <= fquit) // If we reached the user's
goto FINISH ; // limit, we can quit
}
if ((user_quit = user_pressed_escape ()) != 0)
goto FINISH ;
} // Loop: for all initial iters
avg_func /= niters ; // Mean of all points around avgnet
new_fac = 1.0 / niters ; // Weight of each point
sprintf ( msg , " avg=%.6lf best=%.6lf", avg_func, bestfval ) ;
progress_message ( msg ) ;
if (use_grad) { // Also need gradient?
gradlen = 0.0 ; // Will cumulate grad length
for (i=0 ; i<n_grad ; i++) { // Find gradient mean and length
avg_grad[i] /= niters ;
gradlen += avg_grad[i] * avg_grad[i] ;
}
gradlen = sqrt ( gradlen ) ;
grad_weight = 0.5 ;
}
/*
This is the temperature reduction loop and the iteration within
temperature loop.
*/
temp = starttemp ;
tempmult = exp( log( stoptemp / starttemp ) / (ntemps-1)) ;
user_quit = 0 ; // Flags user pressed ESCape
for (itemp=0 ; itemp<ntemps ; itemp++) { // Temp reduction loop
n_good = n_bad = 0 ; // Counts better and worse
sprintf ( msg , "Temp=%.3lf ", temp ) ;
normal_message ( msg ) ;
for (iter=0 ; iter<niters ; iter++) { // Iters per temp loop
if ((n_bad >= 10) &&
((double) n_good / (double) (n_good+n_bad) < 0.15))
break ;
perturb ( avgnet , this , temp ,
reg , density ) ; // Randomly perturb about center
if (use_grad) // Bias per gradient?
weight_used = shift ( grad , this , grad_weight , reg ) ;
if (reg) { // If using regression, estimate
fval = regress ( tptr , sptr ) ; // out weights now
if ((user_quit = user_pressed_escape ()) != 0)
break ;
if (fval >= avg_func) { // If this would raise mean
++n_bad ; // Count this bad point for user
continue ; // Skip it and try again
}
}
if (use_grad) // Need gradient, fval redundant
fval = gradient ( tptr , work1 , work2 , grad ) ;
else if (! reg) // If reg we got fval from regress
fval = trial_error ( tptr ) ;
if ((user_quit = user_pressed_escape ()) != 0)
break ;
if (fval >= avg_func) { // If this would raise mean
++n_bad ; // Count this bad point for user
continue ; // Skip it and try again
}
++n_good ;
if (fval < bestfval) { // If this iteration improved
bestfval = fval ; // then update the best so far
copy_weights ( bestnet , this ) ; // Keep the network
if (bestfval <= fquit) // If we reached the user's
break ; // limit, we can quit
iter -= setback ; // It often pays to keep going
if (iter < 0) // at this temperature if we
iter = 0 ; // are still improving
}
adjust ( avgnet , this , reg , new_fac ) ; // Move center slightly
avg_func = new_fac * fval + (1.0 - new_fac) * avg_func ;
if (use_grad) {
grad_weight = new_fac * weight_used + (1.0 - new_fac) * grad_weight ;
for (i=0 ; i<n_grad ; i++) // Adjust mean gradient
avg_grad[i] = new_fac * grad[i] + (1.0 - new_fac) * avg_grad[i] ;
}
} // Loop: for all iters at a temp
/*
Iters within temp loop now complete
*/
sprintf ( msg , " %.3lf%% improved avg=%.5lf best=%.5lf",
100.0 * n_good / (double) (n_good+n_bad), avg_func, bestfval ) ;
progress_message ( msg ) ;
if (use_grad) {
gradlen = 0.0 ; // Will cumulate grad length
for (i=0 ; i<n_grad ; i++) // Find gradient length
gradlen += avg_grad[i] * avg_grad[i] ;
gradlen = sqrt ( gradlen ) ;
sprintf ( msg , " grad=%.5lf", gradlen ) ;
progress_message ( msg ) ;
}
if (bestfval <= fquit) // If we reached the user's
break ; // limit, we can quit
if (user_quit)
break ;
temp *= tempmult ; // Reduce temp for next pass
} // through this temperature loop
/*
The trials left this weight set and neterr in random condition.
Make them equal to the best, which will be the original
if we never improved.
*/
FINISH:
copy_weights ( this , bestnet ) ; // Return best weights in this net
neterr = bestfval ; // Trials destroyed weights, err
if (reg) {
MEMTEXT ( "SSG: delete SingularValueDecomp" ) ;
delete sptr ;
}
if (user_quit)
return 1 ;
else
return 0 ;
}
float LayerNet::find_grad (
TrainingSet *tptr ,
float *hid2delta ,
float *outdelta ,
float *grad
)
{
int i, j, size, tset, tclass, n, nprev, nnext ;
float error, *dptr, diff, delta, *hid1grad, *hid2grad, *outgrad ;
float *outprev, *prevact, *nextcoefs, *nextdelta, *gradptr ;
char msg[80];
/*
Compute size of each training sample
*/
if (outmod == OUTMOD_CLASSIFY)
size = nin + 1 ;
else if (outmod == OUTMOD_AUTO)
size = nin ;
else if (outmod == OUTMOD_GENERAL)
size = nin + nout ;
/*
Compute length of grad vector and gradient positions in it.
Also point to layer previous to output and its size.
Ditto for layer after hid1.
*/
if (nhid1 == 0) { // No hidden layer
n = nout * (nin+1) ;
outgrad = grad ;
nprev = nin ;
}
else if (nhid2 == 0) { // One hidden layer
n = nhid1 * (nin+1) + nout * (nhid1+1) ;
hid1grad = grad ;
outgrad = grad + nhid1 * (nin+1) ;
outprev = hid1 ;
nprev = nhid1 ;
nnext = nout ;
nextcoefs = out_coefs ;
nextdelta = outdelta ;
}
else { // Two hidden layers
n = nhid1 * (nin+1) + nhid2 * (nhid1+1) + nout * (nhid2+1) ;
hid1grad = grad ;
hid2grad = grad + nhid1 * (nin+1) ;
outgrad = hid2grad + nhid2 * (nhid1+1) ;
outprev = hid2 ;
nprev = nhid2 ;
nnext = nhid2 ;
nextcoefs = hid2_coefs ;
nextdelta = hid2delta ;
}
for (i=0 ; i<n ; i++) // Zero gradient for summing
grad[i] = 0.0 ;
error = 0.0 ; // Will cumulate total error here
for (tset=0 ; tset<tptr->ntrain ; tset++) { // Do all samples
sprintf ( msg , "Learning Pattern Nø %d ", tset) ;
normal_message ( msg ) ;
dptr = tptr->data + size * tset ; // Point to this sample
trial ( dptr ) ; // Evaluate network for it
if (outmod == OUTMOD_AUTO) { // If this is AUTOASSOCIATIVE
for (i=0 ; i<nout ; i++) { // then the expected outputs
diff = *dptr++ - out[i] ; // are just the inputs
error += diff * diff ;
outdelta[i] = diff * actderiv ( out[i] ) ;
}
}
else if (outmod == OUTMOD_CLASSIFY) { // If this is Classification
tclass = (int) dptr[nin] - 1 ; // class is stored after inputs
for (i=0 ; i<nout ; i++) { // Recall that train added a
if (tclass == i) // fraction so that the above
diff = NEURON_ON - out[i] ; // truncation to get tclass is
else // always safe in any radix
diff = NEURON_OFF - out[i] ;
error += diff * diff ;
outdelta[i] = diff * actderiv ( out[i] ) ;
}
}
else if (outmod == OUTMOD_GENERAL) { // If this is GENERAL output
dptr += nin ; // outputs stored after inputs
for (i=0 ; i<nout ; i++) {
diff = *dptr++ - out[i] ;
error += diff * diff ;
outdelta[i] = diff * actderiv ( out[i] ) ;
}
}
/*
Cumulate output gradient
*/
if (nhid1 == 0) // No hidden layer
prevact = tptr->data + size * tset ;
else
prevact = outprev ; // Point to previous layer
gradptr = outgrad ;
for (i=0 ; i<nout ; i++) {
delta = outdelta[i] ;
for (j=0 ; j<nprev ; j++)
*gradptr++ += delta * prevact[j] ;
*gradptr++ += delta ; // Bias activation is always 1
}
/*
Cumulate hid2 gradient (if it exists)
*/
if (nhid2) {
gradptr = hid2grad ;
for (i=0 ; i<nhid2 ; i++) {
delta = 0.0 ;
for (j=0 ; j<nout ; j++)
delta += outdelta[j] * out_coefs[j*(nhid2+1)+i] ;
delta *= actderiv ( hid2[i] ) ;
hid2delta[i] = delta ;
for (j=0 ; j<nhid1 ; j++)
*gradptr++ += delta * hid1[j] ;
*gradptr++ += delta ; // Bias activation is always 1
}
}
/*
Cumulate hid1 gradient (if it exists)
*/
if (nhid1) {
prevact = tptr->data + size * tset ;
gradptr = hid1grad ;
for (i=0 ; i<nhid1 ; i++) {
delta = 0.0 ;
for (j=0 ; j<nnext ; j++)
delta += nextdelta[j] * nextcoefs[j*(nhid1+1)+i] ;
delta *= actderiv ( hid1[i] ) ;
for (j=0 ; j<nin ; j++)
*gradptr++ += delta * prevact[j] ;
*gradptr++ += delta ; // Bias activation is always 1
}
}
} // for all tsets
return error / ((float) tptr->ntrain * (float) nout) ;
}
float LayerNet::conjgrad (
TrainingSet *tptr , // Training set to use
int maxits , // Maximum iterations allowed
float reltol , // Relative error change tolerance
float errtol // Quit if error drops this low
)
{
int i, j, n, iter, pnum, key, retry, max_retry ;
float gam, *g, *h, *outdelta, *hid2delta, *grad, *base ;
float corr, error, *cptr, *gptr, *pptr, maxgrad ;
float prev_err ;
char msg[80];
max_retry = 5 ;
/*
Allocate work memory
*/
MEMTEXT ( "CONJGRAD work" ) ;
if (nhid2) {
hid2delta = (float *) MALLOC ( nhid2 * sizeof(float) ) ;
if (hid2delta == NULL)
return -2.0 ;
}
else
hid2delta = NULL ;
outdelta = (float *) MALLOC ( nout * sizeof(float) ) ;
if (nhid1 == 0) // No hidden layer
n = nout * (nin+1) ;
else if (nhid2 == 0) // One hidden layer
n = nhid1 * (nin+1) + nout * (nhid1+1) ;
else // Two hidden layers
n = nhid1 * (nin+1) + nhid2 * (nhid1+1) + nout * (nhid2+1) ;
grad = (float *) MALLOC ( n * sizeof(float) ) ;
base = (float *) MALLOC ( n * sizeof(float) ) ;
g = (float *) MALLOC ( n * sizeof(float) ) ;
h = (float *) MALLOC ( n * sizeof(float) ) ;
if ((outdelta == NULL) || (grad == NULL) ||
(base == NULL) || (g == NULL) || (h == NULL)) {
if (hid2delta != NULL)
FREE ( hid2delta ) ;
if (outdelta != NULL)
FREE ( outdelta ) ;
if (grad != NULL)
FREE ( grad ) ;
if (base != NULL)
FREE ( base ) ;
if (g != NULL)
FREE ( g ) ;
if (h != NULL)
FREE ( h ) ;
return -2.0 ; // Flags error
}
prev_err = 1.e30 ;
error = find_grad ( tptr , hid2delta , outdelta , grad ) ;
memcpy ( g , grad , n * sizeof(float) ) ;
memcpy ( h , grad , n * sizeof(float) ) ;
/*
Main iteration loop is here
*/
for (iter=0 ; iter<maxits ; iter++) { // Each iter is an epoch
/*
Check current error against user's max. Abort if user pressed ESCape
*/
sprintf ( msg , "Gradient Finding...Iter Nø %d : Error = %lf %%", iter, 100.0 * error ) ;
normal_message ( msg ) ;
if (error <= errtol) // If our error is within user's limit
break ; // then we are done!
if (error <= reltol) // Generally not necessary: reltol<errtol in
break ; // practice, but help silly users
if (kbhit()) { // Was a key pressed?
key = getch () ; // Read it if so
while (kbhit()) // Flush key buffer in case function key
getch () ; // or key was held down
if (key == 27) { // ESCape
error = -error ; // Flags user that ESCape was pressed
break ;
}
}
prev_err = error ;
error = direcmin ( tptr , error , 10 , 1.e-10 ,
0.5 , base , grad ) ;
if (error < 0.0) // Indicates user pressed ESCape
goto CGFINISH ;
if ((2.0 * (prev_err - error)) <= // If this direc gave poor result
(reltol * (prev_err + error + 1.e-10))) { // will use random direc
prev_err = error ; // But first exhaust grad
error = find_grad ( tptr , hid2delta , outdelta , grad ) ;
error = direcmin ( tptr , error , 15 , 1.e-10 ,
1.e-3 , base , grad ) ;
for (retry=0 ; retry<max_retry ; retry++) {
for (i=0 ; i<n ; i++)
grad[i] = (float) (rand() - RANDMAX/2) / (RANDMAX * 10.0) ;
error = direcmin ( tptr , error , 10 , 1.e-10 ,
1.e-2 , base , grad ) ;
if (error < 0.0) // Indicates user pressed ESCape
goto CGFINISH ;
if (retry < max_retry/2)
continue ;
if ((2.0 * (prev_err - error)) >
(reltol * (prev_err + error + 1.e-10)))
break ; // Get out of retry loop if we improved enough
} // For retry
if (retry == max_retry) // If we exhausted all tries
break ; // probably hopeless
memcpy ( g , grad , n * sizeof(float) ) ;
memcpy ( h , grad , n * sizeof(float) ) ;
} // If this dir gave poor result
prev_err = error ;
/*
Setup for next iteration
*/
error = find_grad ( tptr , hid2delta , outdelta , grad ) ;
gam = gamma ( g , grad ) ;
if (gam < 0.0)
gam = 0.0 ;
if (gam > 1.0)
gam = 1.0 ;
find_new_dir ( gam , g , h , grad ) ;
} // This is the end of the main iteration loop
/*
Free work memory
*/
CGFINISH:
MEMTEXT ( "CONJGRAD work" ) ;
if (hid2delta != NULL)
FREE ( hid2delta ) ;
FREE ( outdelta ) ;
FREE ( grad ) ;
FREE ( base ) ;
FREE ( g ) ;
FREE ( h ) ;
return error ;
}
int main (
int argc , // Number of command line arguments (includes prog name)
char *argv[] // Arguments (prog name is argv[0])
)
{
/*
Declarations of local variables
*/
/*
User's command control line related variables are here.
Control_file_number and control_files permit nesting of 'CONTROL' commands.
If control_file_number equals -1, control commands are read from stdin.
Otherwise they are read from that file in FILE *control_files.
Up to MAX_CONTROL_FILES can be stacked.
*/
int control_file_number = -1 ; // Stack pointer for control files
FILE *control_files[MAX_CONTROL_FILES] ; // This is the stack
char *control_line ; // User's commands here
char *command, *rest ; // Pointers to its command and parameter parts
int n_command, n_rest ; // Lengths of those parts
/*
These are network parameters which may be set by the user via commands.
They are initialized to defaults which indicate that the user has not
yet set them. As they are set, their current values are placed here.
When learning is done for a network, their values are copied from here
into the network object. When a network is read, the object's values
are copied from it to here. Otherwise, these variables are not used;
the values in the network object itself are used. The only purpose of
these variables is to keep track of current values.
*/
int net_model = -1 ; // Network model (see NETMOD_? in CONST.H)
int out_model = -1 ; // Output model (see OUTMOD_? in CONST.H)
int n_inputs = -1 ; // Number of input neurons
int n_outputs = -1 ; // Number of output neurons
int n_hidden1 = -1 ; // Number of hidden layer one neurons
int n_hidden2 = -1 ; // Ditto layer 2 (0 if just one hidden layer)
TrainingSet *tset = NULL ; // Training set here
Network *network = NULL ; // Network here
struct LearnParams learn_params ; // General learning parameters
struct AnnealParams anneal_params ; // Simulated annealing parameters
struct GenInitParams geninit_params ; // Genetic initialization parameters
struct KohParams koh_params ; // Kohonen parameters
int classif_output = -1 ; // Current class (0=reject) for classif training
char out_file[80] = "" ; // File for EXECUTE output
double threshold ; // CLASSIFY confusion reject cutoff
/*
Miscellaneous variables
*/
int i, n, m ;
double p ;
char *msg ;
FILE *fp ;
/*
--------------------------------------------------------------------------------
Program starts here.
Verify that a careless user didn't fail to set the integer size
correctly when compiling.
--------------------------------------------------------------------------------
*/
#if VERSION_16_BIT
if (sizeof(int) > 2) {
printf ( "\nRecompile with VERSION_16_BIT set to 0 in CONST.H" ) ;
exit ( 1 ) ;
}
#else
if (sizeof(int) < 4) {
printf ( "\nRecompile with VERSION_16_BIT set to 1 in CONST.H" ) ;
exit ( 1 ) ;
}
#endif
printf ( "\nNEURAL - Program to train and test neural networks" ) ;
printf("\nCopyright (c) 1993 by Academic Press, Inc.");
printf("\nAll rights reserved. Permission is hereby granted, until further notice,");
printf("\nto make copies of this diskette, which are not for resale, provided these");
printf("\ncopies are made from this master diskette only, and provided that the");
printf("\nfollowing copyright notice appears on the diskette label:");
printf("\n(c) 1993 by Academic Press, Inc.");
printf("\nExcept as previously stated, no part of the computer program embodied in");
printf("\nthis diskette may be reproduced or transmitted in any form or by any means,");
printf("\nelectronic or mechanical, including input into storage in any information");
printf("\nsystem for resale, without permission in writing from the publisher.");
printf("\nProduced in the United States of America.");
printf("\nISBN 0-12-479041-0");
/*
Process command line parameters
*/
mem_name[0] = 0 ; // Default is no memory allocation file
for (i=1 ; i<argc ; i++) { // Process all command line args
str_to_upr ( argv[i] ) ; // Easier if all upper case
if (! strcmp ( argv[i] , "/DEBUG" )) {
sscanf ( argv[++i] , "%s" , mem_name ) ;
if ((strlen ( mem_name ) > 1) || ! isalpha ( mem_name[0] )) {
printf ( "\nIllegal DEBUG drive (%s); must be 1 letter." ) ;
exit ( 1 ) ;
}
continue ;
}
printf ( "\nIllegal command line parameter (%s)", argv[i] ) ;
exit ( 1 ) ;
}
/*
Initialize memory allocation debugging
*/
if (strlen ( mem_name )) {
strcat ( mem_name , ":mem.log" ) ;
fp = fopen ( mem_name , "wt" ) ;
if (fp == NULL) {
printf ( "\nCannot open debugging file %s", mem_name ) ;
exit ( 1 ) ;
}
fclose ( fp ) ;
mem_log = 1 ;
}
else
mem_log = 0 ;
mem_used = 0 ;
/*
Initialize defaults
*/
learn_params.init = -1 ;
learn_params.quit_err = 0.0 ;
learn_params.retries = 32767 ;
anneal_params.temps0 = 3 ;
anneal_params.temps = 4 ;
anneal_params.iters0 = 50 ;
anneal_params.iters = 20 ;
anneal_params.setback0 = 50 ;
anneal_params.setback = 20 ;
anneal_params.start0 = 3.0 ;
anneal_params.start = 4.0 ;
anneal_params.stop0 = 1.0 ;
anneal_params.stop = 0.02 ;
geninit_params.pool = 50 ;
geninit_params.gens = 3 ;
geninit_params.climb = 0 ;
geninit_params.overinit = 1.5 ;
geninit_params.pcross = 0.8 ;
geninit_params.pmutate = 0.0001 ;
koh_params.normalization = 0 ; // 0=multiplicative, 1=Z
koh_params.learn_method = 1 ; // 0=additive, 1=subtractive
koh_params.rate = 0.4 ; // learning rate
koh_params.reduction = 0.99 ; // learning rate reduction
learn_params.ap = &anneal_params ;
learn_params.gp = &geninit_params ;
learn_params.kp = &koh_params ;
act_func_init () ; // Initialize interpolation table for activation function
MEMTEXT ( "NEURAL: control_line, msg" ) ;
if (((control_line = (char *) MALLOC ( CONTROL_LINE_LENGTH+1 )) == NULL)
|| ((msg = (char *) MALLOC ( CONTROL_LINE_LENGTH+1 )) == NULL)) {
printf ( "\nInsufficient memory" ) ;
exit ( 1 ) ;
}
/*
Main loop processes all commands
*/
for (;;) {
get_control_line ( control_line , &control_file_number, control_files ) ;
split_control_line ( control_line , &command , &n_command ,
&rest , &n_rest ) ;
if (! n_command) {
if (n_rest) {
sprintf ( msg , "No colon after command: %s", rest ) ;
error_message ( msg ) ;
}
continue ;
}
sprintf ( msg , "%s : %s", command, rest ) ;
normal_message ( msg ) ;
/*
Act on the command
*/
if (! strcmp ( command , "QUIT" ))
break ;
if (! strcmp ( command , "CONTROL" )) {
stack_control_file ( rest , &control_file_number , control_files ) ;
continue ;
}
if (! strcmp ( command , "NETWORK MODEL" )) {
if (! strcmp ( rest , "LAYER" ))
n = NETMOD_LAYER ;
else if (! strcmp ( rest , "KOHONEN" ))
n = NETMOD_KOH ;
else {
sprintf ( msg , "Illegal NETWORK MODEL: %s", rest ) ;
error_message ( msg ) ;
continue ;
}
if (net_model == n)
continue ;
if (ok_to_clear_weights( &network )) {
net_model = n ;
learn_params.init = -1 ;
}
else
warning_message ( "Command aborted" ) ;
continue ;
}
if (! strcmp ( command , "OUTPUT MODEL" )) {
if (! strcmp ( rest , "CLASSIFY" ))
n = OUTMOD_CLASSIFY ;
else if (! strcmp ( rest , "AUTO" ))
n = OUTMOD_AUTO ;
else if (! strcmp ( rest , "GENERAL" ))
n = OUTMOD_GENERAL ;
else {
sprintf ( msg , "Illegal OUTPUT MODEL: %s", rest ) ;
error_message ( msg ) ;
continue ;
}
if (out_model == n)
continue ;
if ((ok_to_clear_tset( &tset )) && (ok_to_clear_weights( &network)))
out_model = n ;
else
warning_message ( "Command aborted" ) ;
continue ;
}
if (! strcmp ( command , "N INPUTS" )) {
m = sscanf ( rest , "%d" , &n ) ;
if ((m <= 0) || (n <= 0) || (n > MAX_INPUTS)) {
sprintf ( msg , "Illegal N INPUTS: %s", rest ) ;
error_message ( msg ) ;
continue ;
}
if (n_inputs == n)
continue ;
if ((ok_to_clear_tset( &tset)) && (ok_to_clear_weights(&network)))
n_inputs = n ;
else
warning_message ( "Command aborted" ) ;
continue ;
}
if (! strcmp ( command , "N OUTPUTS" )) {
m = sscanf ( rest , "%d" , &n ) ;
if ((m <= 0) || (n <= 0) || (n > MAX_OUTPUTS)) {
sprintf ( msg , "Illegal N OUTPUTS: %s", rest ) ;
error_message ( msg ) ;
continue ;
}
if (n_outputs == n)
continue ;
if ((ok_to_clear_tset( &tset)) && (ok_to_clear_weights(&network)))
n_outputs = n ;
else
warning_message ( "Command aborted" ) ;
continue ;
}
if (! strcmp ( command , "N HIDDEN1" )) {
m = sscanf ( rest , "%d" , &n ) ;
if ((m <= 0) || (n < 0) || (n > MAX_HIDDEN)) {
sprintf ( msg , "Illegal N HIDDEN1: %s", rest ) ;
error_message ( msg ) ;
continue ;
}
if (n_hidden1 == n)
continue ;
if (ok_to_clear_weights( &network ))
n_hidden1 = n ;
else
warning_message ( "Command aborted" ) ;
continue ;
}
if (! strcmp ( command , "N HIDDEN2" )) {
m = sscanf ( rest , "%d" , &n ) ;
if ((m <= 0) || (n < 0) || (n > MAX_HIDDEN)) {
sprintf ( msg , "Illegal N HIDDEN2: %s", rest ) ;
error_message ( msg ) ;
continue ;
}
if (n && ! n_hidden1) {
error_message ( "N HIDDEN2 must be 0 if N HIDDEN1 IS 0." ) ;
continue ;
}
if (n_hidden2 == n)
continue ;
if (ok_to_clear_weights( &network ))
n_hidden2 = n ;
else
warning_message ( "Command aborted" ) ;
continue ;
}
if (! strcmp ( command , "TRAIN" )) {
if ((out_model == OUTMOD_AUTO) && (n_outputs != n_inputs)) {
warning_message ( "Setting N OUTPUTS = N INPUTS" ) ;
n_outputs = n_inputs ;
}
if (out_model <= 0)
error_message ( "TRAIN used before OUTPUT MODEL set." ) ;
else if (n_inputs <= 0)
error_message ( "TRAIN used before N INPUTS set." ) ;
else if (n_outputs <= 0)
error_message ( "TRAIN used before N OUTPUTS set." ) ;
else if ((net_model != NETMOD_KOH) && (out_model == OUTMOD_CLASSIFY)
&& (classif_output < 0))
error_message( "CLASSIFY output mode but CLASSIFY OUTPUT not set.");
else if ((net_model == NETMOD_KOH) && (out_model != OUTMOD_CLASSIFY))
error_message( "KOHONEN network requires CLASSIFY output mode.");
else {
if (tset == NULL) {
MEMTEXT ( "NEURAL: new tset" ) ;
tset = new TrainingSet ( out_model , n_inputs , n_outputs ) ;
}
tset->train ( rest , classif_output ) ;
}
continue ;
}
if (check_anneal ( command , rest , &anneal_params ))
continue ;
if (check_genetic ( command , rest , &geninit_params ))
continue ;
if (check_kohonen ( command , rest , &koh_params , &network ))
continue ;
if (check_learn_params ( command , rest , &learn_params , net_model ))
continue ;
if (! strcmp ( command , "LEARN" )) {
if ((tset == NULL) || (tset->ntrain == 0)) {
error_message ( "Cannot LEARN; No training set exists." ) ;
continue ;
}
if ((net_model == NETMOD_KOH) && (out_model != OUTMOD_CLASSIFY)) {
error_message( "KOHONEN network requires CLASSIFY output mode.");
continue ;
}
if (learn_params.init < 0) {
error_message( "Initialization method not set.");
continue ;
}
if (network == NULL) {
if (net_model == NETMOD_LAYER) {
if (n_hidden1 < 0) {
error_message ( "LEARN used before N HIDDEN1 set." ) ;
continue ;
}
else if (n_hidden2 < 0) {
error_message ( "LEARN used before N HIDDEN2 set." ) ;
continue ;
}
else {
MEMTEXT ( "NEURAL: new LayerNet" ) ;
network = new LayerNet ( out_model , n_inputs , n_hidden1 ,
n_hidden2 , n_outputs , 1 , 1 ) ;
}
}
else if (net_model == NETMOD_KOH) {
MEMTEXT ( "NEURAL: new KohNet" ) ;
network = new KohNet ( n_inputs , n_outputs ,
&koh_params , 1 , 1 ) ;
}
}
if ((network == NULL) || (! network->ok)) { // Malloc failure?
memory_message ( "to create network." ) ;
if (network != NULL) {
delete network ;
network = NULL ;
}
continue ;
}
network->learn ( tset , &learn_params ) ;
if (network->neterr > 0.999999) { // Indicates massive failure
MEMTEXT ( "NEURAL: learn failure delete network" ) ;
delete network ;
network = NULL ;
}
else {
sprintf ( msg , "Final error = %.4lf%% of max possible",
100.0 * network->neterr ) ;
normal_message ( msg ) ;
}
continue ;
}
if (! strcmp ( command , "SAVE WEIGHTS" )) {
if (network == NULL)
error_message ( "There are no learned weights to save." ) ;
else
wt_save ( network , net_model , 0 , rest ) ;
continue ;
}
if (! strcmp ( command , "RESTORE WEIGHTS" )) {
if (network != NULL) {
MEMTEXT ( "NEURAL: delete network for restore" ) ;
delete network ;
network = NULL ;
}
network = wt_restore ( rest , &net_model ) ;
if (network == NULL)
continue ;
if (tset != NULL) {
if ((tset->nin != network->nin)
|| (tset->nout != network->nout)
|| (tset->outmod != network->outmod)) {
error_message ( "Network conflicts with existing training set.");
continue ;
}
}
out_model = network->outmod ;
n_inputs = network->nin ;
n_outputs = network->nout ;
if (net_model == NETMOD_LAYER) {
n_hidden1 = ((LayerNet*) network)->nhid1 ;
n_hidden2 = ((LayerNet*) network)->nhid2 ;
}
if (net_model == NETMOD_KOH)
koh_params.normalization = ((KohNet *) network)->normalization ;
learn_params.init = -1 ;
continue ;
}
if (! strcmp ( command , "CLEAR TRAINING" )) {
if (tset != NULL) {
MEMTEXT ( "NEURAL: delete tset" ) ;
delete tset ;
tset = NULL ;
}
continue ;
}
if (! strcmp ( command , "CLEAR WEIGHTS" )) {
if (network != NULL) {
MEMTEXT ( "NEURAL: delete network" ) ;
delete network ;
network = NULL ;
}
continue ;
}
if (! strcmp ( command , "CLASSIFY OUTPUT" )) {
if (net_model == NETMOD_KOH) {
error_message ( "Cannot specify output for KOHONEN model." ) ;
continue ;
}
if (n_outputs < 0) {
error_message ( "CLASSIFY OUTPUT used before N OUTPUTS set." ) ;
continue ;
}
if (out_model != OUTMOD_CLASSIFY) {
error_message
( "CLASSIFY OUTPUT only valid when OUTPUT MODEL:CLASSIFY" ) ;
continue ;
}
m = sscanf ( rest , "%d" , &n ) ;
if ((m <= 0) || (n < 0)) {
sprintf ( msg , "Illegal CLASSIFY OUTPUT: %s", rest ) ;
error_message ( msg ) ;
}
else if (n > n_outputs) {
sprintf ( msg , "CLASSIFY OUTPUT (%d) exceeds N OUTPUTS (%d)",
n, n_outputs ) ;
error_message ( msg ) ;
}
else
classif_output = n ;
continue ;
}
if (! strcmp ( command , "OUTPUT FILE" )) {
strcpy ( out_file , rest ) ;
continue ;
}
if (! strcmp ( command , "EXECUTE" )) {
if (network == NULL)
error_message ( "There is no trained network" ) ;
else
network->execute_from_file ( rest , out_file ) ;
continue ;
}
if (! strcmp ( command , "CLASSIFY" )) {
if (network == NULL)
error_message ( "There is no trained network" ) ;
else if (out_model != OUTMOD_CLASSIFY)
error_message ( "CLASSIFY valid only in CLASSIFY output mode" ) ;
else
network->classify_from_file ( rest , threshold ) ;
continue ;
}
if (! strcmp ( command , "RESET CONFUSION" )) {
if (network == NULL)
error_message ( "There is no trained network" ) ;
else
network->reset_confusion () ;
continue ;
}
if (! strcmp ( command , "CONFUSION THRESHOLD" )) {
p = atof ( rest ) ;
if ((p < 0.0) || (p > 100.0)) {
sprintf ( msg , "Illegal CONFUSION THRESHOLD: %s", rest ) ;
error_message ( msg ) ;
}
else
threshold = p / 100.0 ;
continue ;
}
if (! strcmp ( command , "SHOW CONFUSION" )) {
if (network == NULL)
error_message ( "There is no trained network" ) ;
else if (out_model != OUTMOD_CLASSIFY)
error_message ( "CONFUSION valid only in CLASSIFY output mode" ) ;
else
network->show_confusion () ;
continue ;
}
if (! strcmp ( command , "SAVE CONFUSION" )) {
if (network == NULL)
error_message ( "There is no trained network" ) ;
else if (out_model != OUTMOD_CLASSIFY)
error_message ( "CONFUSION valid only in CLASSIFY output mode" ) ;
else
network->save_confusion ( rest ) ;
continue ;
}
sprintf ( msg , "Unknown command: %s", command ) ;
error_message ( msg ) ;
} // Endless command loop
MEMTEXT ( "NEURAL: control_line, msg" ) ;
FREE ( control_line ) ;
FREE ( msg ) ;
MEMCLOSE () ;
return 0 ;
}
