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);
}
/* INTERNAL FUNCTION
The SARprop- algorithm
*/
void fann_update_weights_sarprop(struct fann *ann, unsigned int epoch, unsigned int first_weight, unsigned int past_end)
{
fann_type *train_slopes = ann->train_slopes;
fann_type *weights = ann->weights;
fann_type *prev_steps = ann->prev_steps;
fann_type *prev_train_slopes = ann->prev_train_slopes;
fann_type prev_step, slope, prev_slope, next_step = 0, same_sign;
/* These should be set from variables */
float increase_factor = ann->rprop_increase_factor; /*1.2; */
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) */
float delta_min = 0.000001f;
float delta_max = ann->rprop_delta_max; /*50.0; */
float weight_decay_shift = ann->sarprop_weight_decay_shift; /* ld 0.01 = -6.644 */
float step_error_threshold_factor = ann->sarprop_step_error_threshold_factor; /* 0.1 */
float step_error_shift = ann->sarprop_step_error_shift; /* ld 3 = 1.585 */
float T = ann->sarprop_temperature;
float MSE = fann_get_MSE(ann);
float RMSE = (float)sqrt(MSE);
unsigned int i = first_weight;
/* for all weights; TODO: are biases included? */
for(; i != past_end; i++)
{
/* TODO: confirm whether 1x10^-6 == delta_min is really better */
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)*/
slope = -train_slopes[i] - weights[i] * (fann_type)fann_exp2(-T * epoch + weight_decay_shift);
/* TODO: is prev_train_slopes[i] 0.0 in the beginning? */
prev_slope = prev_train_slopes[i];
same_sign = prev_slope * slope;
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 (slope < 0.0)
weights[i] += next_step;
else
weights[i] -= next_step;
}
else if(same_sign < 0.0)
{
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);
slope = 0.0;
}
else
{
if(slope < 0.0)
weights[i] += prev_step;
else
weights[i] -= prev_step;
}
/*if(i == 2){
* printf("weight=%f, slope=%f, next_step=%f, prev_step=%f\n", weights[i], slope, next_step, prev_step);
* } */
/* update global data arrays */
prev_steps[i] = next_step;
prev_train_slopes[i] = slope;
train_slopes[i] = 0.0;
}
}