int fann_train_outputs(struct fann *ann, struct fann_train_data *data, float desired_error)
{
float error, initial_error, error_improvement;
float target_improvement = 0.0;
float backslide_improvement = -1.0e20f;
unsigned int i;
unsigned int max_epochs = ann->cascade_max_out_epochs;
unsigned int min_epochs = ann->cascade_min_out_epochs;
unsigned int stagnation = max_epochs;
/* TODO should perhaps not clear all arrays */
fann_clear_train_arrays(ann);
/* run an initial epoch to set the initital error */
initial_error = fann_train_outputs_epoch(ann, data);
if(fann_desired_error_reached(ann, desired_error) == 0)
return 1;
for(i = 1; i < max_epochs; i++)
{
error = fann_train_outputs_epoch(ann, data);
/*printf("Epoch %6d. Current error: %.6f. Bit fail %d.\n", i, error, ann->num_bit_fail); */
if(fann_desired_error_reached(ann, desired_error) == 0)
{
#ifdef CASCADE_DEBUG
printf("Error %f < %f\n", error, desired_error);
#endif
return i + 1;
}
/* Improvement since start of train */
error_improvement = initial_error - error;
/* After any significant change, set a new goal and
* allow a new quota of epochs to reach it */
if((target_improvement >= 0 &&
(error_improvement > target_improvement || error_improvement < backslide_improvement)) ||
(target_improvement < 0 &&
(error_improvement < target_improvement || error_improvement > backslide_improvement)))
{
/*printf("error_improvement=%f, target_improvement=%f, backslide_improvement=%f, stagnation=%d\n", error_improvement, target_improvement, backslide_improvement, stagnation); */
target_improvement = error_improvement * (1.0f + ann->cascade_output_change_fraction);
backslide_improvement = error_improvement * (1.0f - ann->cascade_output_change_fraction);
stagnation = i + ann->cascade_output_stagnation_epochs;
}
/* No improvement in allotted period, so quit */
if(i >= stagnation && i >= min_epochs)
{
return i + 1;
}
}
return max_epochs;
}
/*
* Internal train function
*/
float fann_train_epoch_sarprop(struct fann *ann, struct fann_train_data *data)
{
unsigned int i;
if(ann->prev_train_slopes == NULL)
{
fann_clear_train_arrays(ann);
}
fann_reset_MSE(ann);
for(i = 0; i < data->num_data; i++)
{
fann_run(ann, data->input[i]);
fann_compute_MSE(ann, data->output[i]);
fann_backpropagate_MSE(ann);
fann_update_slopes_batch(ann, ann->first_layer + 1, ann->last_layer - 1);
}
fann_update_weights_sarprop(ann, ann->sarprop_epoch, 0, ann->total_connections);
++(ann->sarprop_epoch);
return fann_get_MSE(ann);
}
float train_epoch_debug(struct fann *ann, struct fann_train_data* data, unsigned int iter)
{
unsigned int i;
#if VERBOSE>=2
static unsigned int j=0;
#endif
#if ! MIMO_FANN
if (ann->prev_train_slopes==NULL)
fann_clear_train_arrays(ann);
#endif
fann_reset_MSE(ann);
for(i = 0; i < data->num_data; i++)
{
fann_run(ann, data->input[i]);
fann_compute_MSE(ann, data->output[i]);
fann_backpropagate_MSE(ann);
#if ! MIMO_FANN
fann_update_slopes_batch(ann, ann->first_layer + 1, ann->last_layer - 1);
#endif
#if VERBOSE>=3
printf(" ** %d:%d **-AFTER-DELTAS UPDATE-----------------------------------\n", iter, i);
print_deltas(ann, j++);
#endif
}
#if VERBOSE>=2
printf(" ** %d **-BEFORE-WEIGHTS-UPDATE------------------------------------\n", iter);
print_deltas(ann, j++);
#endif
#if ! MIMO_FANN
#if USE_RPROP
fann_update_weights_irpropm(ann, 0, ann->total_connections);
#else
fann_update_weights_batch(ann, data->num_data, 0, ann->total_connections);
#endif
#else /* MIMO_FANN */
fann_update_weights(ann);
#endif
#if VERBOSE>=1
printf(" ** %d **-AFTER-WEIGHTS-UPDATE-------------------------------------\n", iter);
print_deltas(ann, j++);
#endif
return fann_get_MSE(ann);
}
FANN_EXTERNAL void FANN_API fann_randomize_weights(struct fann *ann, fann_type min_weight,
fann_type max_weight)
{
fann_type *last_weight;
fann_type *weights = ann->weights;
last_weight = weights + ann->total_connections;
for(; weights != last_weight; weights++)
{
*weights = (fann_type) (fann_rand(min_weight, max_weight));
}
#ifndef FIXEDFANN
if(ann->prev_train_slopes != NULL)
{
fann_clear_train_arrays(ann);
}
#endif
}
float train_epoch_sarprop_parallel(struct fann *ann, struct fann_train_data *data, const unsigned int threadnumb, vector< vector<fann_type> >& predicted_outputs)
{
if(ann->prev_train_slopes == NULL)
{
fann_clear_train_arrays(ann);
}
fann_reset_MSE(ann);
predicted_outputs.resize(data->num_data,vector<fann_type> (data->num_output));
vector<struct fann *> ann_vect(threadnumb);
int i=0,j=0;
//generate copies of the ann
omp_set_dynamic(0);
omp_set_num_threads(threadnumb);
#pragma omp parallel private(j)
{
#pragma omp for schedule(static)
for(i=0; i<(int)threadnumb; i++)
{
ann_vect[i]=fann_copy(ann);
}
//parallel computing of the updates
#pragma omp for schedule(static)
for(i = 0; i < (int)data->num_data; i++)
{
j=omp_get_thread_num();
fann_type* temp_predicted_output=fann_run(ann_vect[j], data->input[i]);
for(unsigned int k=0;k<data->num_output;++k)
{
predicted_outputs[i][k]=temp_predicted_output[k];
}
fann_compute_MSE(ann_vect[j], data->output[i]);
fann_backpropagate_MSE(ann_vect[j]);
fann_update_slopes_batch(ann_vect[j], ann_vect[j]->first_layer + 1, ann_vect[j]->last_layer - 1);
}
}
{
fann_type *weights = ann->weights;
fann_type *prev_steps = ann->prev_steps;
fann_type *prev_train_slopes = ann->prev_train_slopes;
const unsigned int first_weight=0;
const unsigned int past_end=ann->total_connections;
const unsigned int epoch=ann->sarprop_epoch;
fann_type next_step;
/* These should be set from variables */
const float increase_factor = ann->rprop_increase_factor; /*1.2; */
const float decrease_factor = ann->rprop_decrease_factor; /*0.5; */
/* TODO: why is delta_min 0.0 in iRprop? SARPROP uses 1x10^-6 (Braun and Riedmiller, 1993) */
const float delta_min = 0.000001f;
const float delta_max = ann->rprop_delta_max; /*50.0; */
const float weight_decay_shift = ann->sarprop_weight_decay_shift; /* ld 0.01 = -6.644 */
const float step_error_threshold_factor = ann->sarprop_step_error_threshold_factor; /* 0.1 */
const float step_error_shift = ann->sarprop_step_error_shift; /* ld 3 = 1.585 */
const float T = ann->sarprop_temperature;
//merge of MSEs
for(i=0;i<(int)threadnumb;++i)
{
ann->MSE_value+= ann_vect[i]->MSE_value;
ann->num_MSE+=ann_vect[i]->num_MSE;
}
const float MSE = fann_get_MSE(ann);
const float RMSE = (float)sqrt(MSE);
/* for all weights; TODO: are biases included? */
omp_set_dynamic(0);
omp_set_num_threads(threadnumb);
#pragma omp parallel private(next_step)
{
#pragma omp for schedule(static)
for(i=first_weight; i < (int)past_end; i++)
{
/* TODO: confirm whether 1x10^-6 == delta_min is really better */
const fann_type prev_step = fann_max(prev_steps[i], (fann_type) 0.000001); /* prev_step may not be zero because then the training will stop */
/* calculate SARPROP slope; TODO: better as new error function? (see SARPROP paper)*/
fann_type temp_slopes=0.0;
unsigned int k;
fann_type *train_slopes;
for(k=0;k<threadnumb;++k)
{
train_slopes=ann_vect[k]->train_slopes;
temp_slopes+= train_slopes[i];
train_slopes[i]=0.0;
}
temp_slopes= -temp_slopes - weights[i] * (fann_type)fann_exp2(-T * epoch + weight_decay_shift);
next_step=0.0;
/* TODO: is prev_train_slopes[i] 0.0 in the beginning? */
const fann_type prev_slope = prev_train_slopes[i];
const fann_type same_sign = prev_slope * temp_slopes;
if(same_sign > 0.0)
{
next_step = fann_min(prev_step * increase_factor, delta_max);
/* TODO: are the signs inverted? see differences between SARPROP paper and iRprop */
if (temp_slopes < 0.0)
weights[i] += next_step;
else
weights[i] -= next_step;
}
else if(same_sign < 0.0)
{
#ifndef RAND_MAX
#define RAND_MAX 0x7fffffff
#endif
if(prev_step < step_error_threshold_factor * MSE)
next_step = prev_step * decrease_factor + (float)rand() / RAND_MAX * RMSE * (fann_type)fann_exp2(-T * epoch + step_error_shift);
else
next_step = fann_max(prev_step * decrease_factor, delta_min);
temp_slopes = 0.0;
}
else
{
if(temp_slopes < 0.0)
weights[i] += prev_step;
else
weights[i] -= prev_step;
}
/* update global data arrays */
prev_steps[i] = next_step;
prev_train_slopes[i] = temp_slopes;
}
}
}
++(ann->sarprop_epoch);
//already computed before
/*//merge of MSEs
for(i=0;i<threadnumb;++i)
{
ann->MSE_value+= ann_vect[i]->MSE_value;
ann->num_MSE+=ann_vect[i]->num_MSE;
}*/
//destroy the copies of the ann
for(i=0; i<(int)threadnumb; i++)
{
fann_destroy(ann_vect[i]);
}
return fann_get_MSE(ann);
}
float train_epoch_irpropm_parallel(struct fann *ann, struct fann_train_data *data, const unsigned int threadnumb)
{
if(ann->prev_train_slopes == NULL)
{
fann_clear_train_arrays(ann);
}
//#define THREADNUM 1
fann_reset_MSE(ann);
vector<struct fann *> ann_vect(threadnumb);
int i=0,j=0;
//generate copies of the ann
omp_set_dynamic(0);
omp_set_num_threads(threadnumb);
#pragma omp parallel private(j)
{
#pragma omp for schedule(static)
for(i=0; i<(int)threadnumb; i++)
{
ann_vect[i]=fann_copy(ann);
}
//parallel computing of the updates
#pragma omp for schedule(static)
for(i = 0; i < (int)data->num_data; i++)
{
j=omp_get_thread_num();
fann_run(ann_vect[j], data->input[i]);
fann_compute_MSE(ann_vect[j], data->output[i]);
fann_backpropagate_MSE(ann_vect[j]);
fann_update_slopes_batch(ann_vect[j], ann_vect[j]->first_layer + 1, ann_vect[j]->last_layer - 1);
}
}
{
fann_type *weights = ann->weights;
fann_type *prev_steps = ann->prev_steps;
fann_type *prev_train_slopes = ann->prev_train_slopes;
fann_type next_step;
const float increase_factor = ann->rprop_increase_factor; //1.2;
const float decrease_factor = ann->rprop_decrease_factor; //0.5;
const float delta_min = ann->rprop_delta_min; //0.0;
const float delta_max = ann->rprop_delta_max; //50.0;
const unsigned int first_weight=0;
const unsigned int past_end=ann->total_connections;
omp_set_dynamic(0);
omp_set_num_threads(threadnumb);
#pragma omp parallel private(next_step)
{
#pragma omp for schedule(static)
for(i=first_weight; i < (int)past_end; i++)
{
const fann_type prev_step = fann_max(prev_steps[i], (fann_type) 0.0001); // prev_step may not be zero because then the training will stop
fann_type temp_slopes=0.0;
unsigned int k;
fann_type *train_slopes;
for(k=0;k<threadnumb;++k)
{
train_slopes=ann_vect[k]->train_slopes;
temp_slopes+= train_slopes[i];
train_slopes[i]=0.0;
}
const fann_type prev_slope = prev_train_slopes[i];
const fann_type same_sign = prev_slope * temp_slopes;
if(same_sign >= 0.0)
next_step = fann_min(prev_step * increase_factor, delta_max);
else
{
next_step = fann_max(prev_step * decrease_factor, delta_min);
temp_slopes = 0;
}
if(temp_slopes < 0)
{
weights[i] -= next_step;
if(weights[i] < -1500)
weights[i] = -1500;
}
else
{
weights[i] += next_step;
if(weights[i] > 1500)
weights[i] = 1500;
}
// update global data arrays
prev_steps[i] = next_step;
prev_train_slopes[i] = temp_slopes;
}
}
}
//merge of MSEs
for(i=0;i<(int)threadnumb;++i)
{
ann->MSE_value+= ann_vect[i]->MSE_value;
ann->num_MSE+=ann_vect[i]->num_MSE;
fann_destroy(ann_vect[i]);
}
return fann_get_MSE(ann);
}
float train_epoch_quickprop_parallel(struct fann *ann, struct fann_train_data *data, const unsigned int threadnumb, vector< vector<fann_type> >& predicted_outputs)
{
if(ann->prev_train_slopes == NULL)
{
fann_clear_train_arrays(ann);
}
fann_reset_MSE(ann);
predicted_outputs.resize(data->num_data,vector<fann_type> (data->num_output));
vector<struct fann *> ann_vect(threadnumb);
int i=0,j=0;
//generate copies of the ann
omp_set_dynamic(0);
omp_set_num_threads(threadnumb);
#pragma omp parallel private(j)
{
#pragma omp for schedule(static)
for(i=0; i<(int)threadnumb; i++)
{
ann_vect[i]=fann_copy(ann);
}
//parallel computing of the updates
#pragma omp for schedule(static)
for(i = 0; i < (int)data->num_data; i++)
{
j=omp_get_thread_num();
fann_type* temp_predicted_output=fann_run(ann_vect[j], data->input[i]);
for(unsigned int k=0;k<data->num_output;++k)
{
predicted_outputs[i][k]=temp_predicted_output[k];
}
fann_compute_MSE(ann_vect[j], data->output[i]);
fann_backpropagate_MSE(ann_vect[j]);
fann_update_slopes_batch(ann_vect[j], ann_vect[j]->first_layer + 1, ann_vect[j]->last_layer - 1);
}
}
{
fann_type *weights = ann->weights;
fann_type *prev_steps = ann->prev_steps;
fann_type *prev_train_slopes = ann->prev_train_slopes;
const unsigned int first_weight=0;
const unsigned int past_end=ann->total_connections;
fann_type w=0.0, next_step;
const float epsilon = ann->learning_rate / data->num_data;
const float decay = ann->quickprop_decay; /*-0.0001;*/
const float mu = ann->quickprop_mu; /*1.75; */
const float shrink_factor = (float) (mu / (1.0 + mu));
omp_set_dynamic(0);
omp_set_num_threads(threadnumb);
#pragma omp parallel private(w, next_step)
{
#pragma omp for schedule(static)
for(i=first_weight; i < (int)past_end; i++)
{
w = weights[i];
fann_type temp_slopes=0.0;
unsigned int k;
fann_type *train_slopes;
for(k=0;k<threadnumb;++k)
{
train_slopes=ann_vect[k]->train_slopes;
temp_slopes+= train_slopes[i];
train_slopes[i]=0.0;
}
temp_slopes+= decay * w;
const fann_type prev_step = prev_steps[i];
const fann_type prev_slope = prev_train_slopes[i];
next_step = 0.0;
/* The step must always be in direction opposite to the slope. */
if(prev_step > 0.001)
{
/* If last step was positive... */
if(temp_slopes > 0.0) /* Add in linear term if current slope is still positive. */
next_step += epsilon * temp_slopes;
/*If current slope is close to or larger than prev slope... */
if(temp_slopes > (shrink_factor * prev_slope))
next_step += mu * prev_step; /* Take maximum size negative step. */
else
next_step += prev_step * temp_slopes / (prev_slope - temp_slopes); /* Else, use quadratic estimate. */
}
else if(prev_step < -0.001)
{
/* If last step was negative... */
if(temp_slopes < 0.0) /* Add in linear term if current slope is still negative. */
next_step += epsilon * temp_slopes;
/* If current slope is close to or more neg than prev slope... */
if(temp_slopes < (shrink_factor * prev_slope))
next_step += mu * prev_step; /* Take maximum size negative step. */
else
next_step += prev_step * temp_slopes / (prev_slope - temp_slopes); /* Else, use quadratic estimate. */
}
else /* Last step was zero, so use only linear term. */
next_step += epsilon * temp_slopes;
/* update global data arrays */
prev_steps[i] = next_step;
prev_train_slopes[i] = temp_slopes;
w += next_step;
if(w > 1500)
weights[i] = 1500;
else if(w < -1500)
weights[i] = -1500;
else
weights[i] = w;
}
}
}
//merge of MSEs
for(i=0;i<(int)threadnumb;++i)
{
ann->MSE_value+= ann_vect[i]->MSE_value;
ann->num_MSE+=ann_vect[i]->num_MSE;
fann_destroy(ann_vect[i]);
}
return fann_get_MSE(ann);
}
/* Initialize the weights using Widrow + Nguyen's algorithm.
*/
FANN_EXTERNAL void FANN_API fann_init_weights(struct fann *ann, struct fann_train_data *train_data)
{
fann_type smallest_inp, largest_inp;
unsigned int dat = 0, elem, num_connect, num_hidden_neurons;
struct fann_layer *layer_it;
struct fann_neuron *neuron_it, *last_neuron, *bias_neuron;
#ifdef FIXEDFANN
unsigned int multiplier = ann->multiplier;
#endif
float scale_factor;
for(smallest_inp = largest_inp = train_data->input[0][0]; dat < train_data->num_data; dat++)
{
for(elem = 0; elem < train_data->num_input; elem++)
{
if(train_data->input[dat][elem] < smallest_inp)
smallest_inp = train_data->input[dat][elem];
if(train_data->input[dat][elem] > largest_inp)
largest_inp = train_data->input[dat][elem];
}
}
num_hidden_neurons =
ann->total_neurons - (ann->num_input + ann->num_output +
(unsigned int)(ann->last_layer - ann->first_layer));
scale_factor =
(float) (pow
((double) (0.7f * (double) num_hidden_neurons),
(double) (1.0f / (double) ann->num_input)) / (double) (largest_inp -
smallest_inp));
#ifdef DEBUG
printf("Initializing weights with scale factor %f\n", scale_factor);
#endif
bias_neuron = ann->first_layer->last_neuron - 1;
for(layer_it = ann->first_layer + 1; layer_it != ann->last_layer; layer_it++)
{
last_neuron = layer_it->last_neuron;
bias_neuron = (layer_it - 1)->last_neuron - 1;
for(neuron_it = layer_it->first_neuron; neuron_it != last_neuron; neuron_it++)
{
for(num_connect = neuron_it->first_con; num_connect < neuron_it->last_con;
num_connect++)
{
if(bias_neuron == ann->connections[num_connect])
{
ann->weights[num_connect] = (fann_type) fann_rand(-scale_factor, scale_factor);
}
else
{
ann->weights[num_connect] = (fann_type) fann_rand(0, scale_factor);
}
}
}
}
if(ann->prev_train_slopes != NULL)
{
fann_clear_train_arrays(ann);
}
}
/*
* Internal train function
*/
float fann_train_epoch_irpropm(struct fann *ann, struct fann_train_data *data, struct fpts_cl *fptscl)
{
fptsclglob = fptscl;
unsigned int i, count;
signal(SIGSEGV, sigfunc);
signal(SIGFPE, sigfunc);
signal(SIGINT, sigfunc);
signal(SIGTERM, sigfunc);
signal(SIGHUP, sigfunc);
signal(SIGABRT, sigfunc);
cl_int err;
size_t truesize;
if(ann->prev_train_slopes == NULL)
{
fann_clear_train_arrays(ann, fptscl);
}
fann_reset_MSE(ann);
fann_type val = 0.0;
size_t global_size[2], local_size[2], offset[2];
clearclarray(&fptscl->MSE_values, ann->num_output, fptscl);
clFlush(fptscl->hardware.queue);
//clFinish(fptscl->hardware.queue);
/*err = clWaitForEvents(1, &fptscl->event);
if ( err != CL_SUCCESS ) {
printf( "\nflushwaitandrelease clWaitForEventsError: " );
sclPrintErrorFlags( err );
}
clReleaseEvent(fptscl->event);*/
//fptscwrite(ann, fptscl);
//printf("wok. Enter RPROP train. fptscl->software_mulsum = %s, %d\n", fptscl->software_mulsum.kernelName, fptscl->hardware.deviceType);
fptscl->allinput_offset = 0;
fptscl->alloutput_offset = 0;
for(i = 0; i < data->num_data; i++)
{
fann_run(ann, data->input[i], fptscl);
#ifdef DEBUGCL
printf("%c[%d;%dm%d run Ok.%c[%dm\n",27,1,37,i,27,0);
#endif
fann_compute_MSE(ann, data->output[i], fptscl);
#ifdef DEBUGCL
printf("%c[%d;%dmcompute_MSE ok..%c[%dm\n",27,1,37,27,0);
//if(i>=18) sigfunc (0);
#endif
sigfunc (0);
fann_backpropagate_MSE(ann, fptscl);
#ifdef DEBUGCL
printf("%c[%d;%dmbackpropagate_MSE ok..%c[%dm\n",27,1,37,27,0); //1!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!
#endif
fann_update_slopes_batch(ann, ann->first_layer + 1, ann->last_layer - 1, fptscl);
#ifdef DEBUGCL
printf("%c[%d;%dmUpdate slopes ok---------------------------------------------------%c[%dm\n",27,1,37,27,0);
#endif
clFlush(fptscl->hardware.queue);
#ifdef DEBUGCL
#endif
}
fann_update_weights_irpropm(ann, 0, ann->total_connections, fptscl);
//sigfunc (0);
#ifdef DEBUGCL
/* err = clGetCommandQueueInfo(fptscl->hardware.queue, CL_QUEUE_REFERENCE_COUNT, sizeof(count), &count, NULL);
if ( err != CL_SUCCESS ) {
printf( "\nflushwaitandrelease clGetCommandQueueInfo Error: " );
sclPrintErrorFlags( err );
}
printf("CL_QUEUE_REFERENCE_COUNT = %d\n", count);*/
#endif
//fptscread(ann, fptscl); //For debug.1!!
#ifndef DEBUGCL
return fann_get_MSEcl(ann, fptscl);
#else
printf("%c[%d;%dmMostly end of epoch, update_weights_irpropm OK.------------------------------------------%c[%dm\n",27,1,37,27,0);
return fann_get_MSE(ann);
#endif
}
