#!/usr/bin/env python
"""
basic_fusion.py
VARPA, University of Coruna
Mondejar Guerra, Victor M.
30 Oct 2017
"""
from train_SVM import *
# Compute the basic rule from the list of probs
# selected by rule index:
# 0 = product
# 1 = sum
# 2 = minimum
# 3 = minimum
# 4 = majority
# and return the predictions
def basic_rules(probs_ensemble, rule_index):
n_ensembles, n_instances, n_classes = probs_ensemble.shape
predictions_rule = np.zeros(n_instances)
# Product rule
if rule_index == 0:
probs_rule = np.ones([n_instances, n_classes])
for p in range(n_instances):
for e in range(n_ensembles):
probs_rule[p] = probs_rule[p] * probs_ensemble[e,p]
predictions_rule[p] = np.argmax(probs_rule[p])
# Sum rule
elif rule_index == 1:
probs_rule = np.zeros([n_instances, n_classes])
for p in range(n_instances):
for e in range(n_ensembles):
probs_rule[p] = probs_rule[p] + probs_ensemble[e,p]
predictions_rule[p] = np.argmax(probs_rule[p])
# Minimum rule
elif rule_index == 2:
probs_rule = np.ones([n_instances, n_classes])
for p in range(n_instances):
for e in range(n_ensembles):
probs_rule[p] = np.minimum(probs_rule[p], probs_ensemble[e,p])
predictions_rule[p] = np.argmax(probs_rule[p])
# Maximum rule
elif rule_index == 3:
probs_rule = np.zeros([n_instances, n_classes])
for p in range(n_instances):
for e in range(n_ensembles):
probs_rule[p] = np.maximum(probs_rule[p], probs_ensemble[e,p])
predictions_rule[p] = np.argmax(probs_rule[p])
# Majority rule
elif rule_index == 4:
rank_rule = np.zeros([n_instances, n_classes])
# Just simply adds the position of the ranking
for p in range(n_instances):
for e in range(n_ensembles):
rank = np.argsort(probs_ensemble[e,p])
for j in range(n_classes):
rank_rule[p,rank[j]] = rank_rule[p,rank[j]] + j
predictions_rule[p] = np.argmax(rank_rule[p])
return predictions_rule
def main():
DS = 'DS2'
print("Runing basic_fusion.py!" + DS)
oversamp = '' #'', 'SMOTEENN/', 'SMOTE/', 'SMOTETomek/', 'ADASYN/'
# Load gt labelso
eval_labels = np.loadtxt('/home/mondejar/Dropbox/ECG/code/ecg_classification/python/mit_db/' + DS + '_labels.csv')
# Configuration
results_path = '/home/mondejar/Dropbox/ECG/code/ecg_classification/python/results/ovo/MLII/'
if DS == 'DS2':
model_RR = results_path + oversamp + 'rm_bsln/' + 'maxRR/' + 'RR/' + 'norm_RR/' + 'weighted/' + 'C_0.001' + '_decision_ovo.csv'
model_wvl = results_path + oversamp + 'rm_bsln/' + 'maxRR/' + 'wvlt/' + 'weighted/' + 'C_0.001' + '_decision_ovo.csv'
model_LBP = results_path + oversamp + 'rm_bsln/' + 'maxRR/' + 'lbp/' + 'weighted/' + 'C_0.001' + '_decision_ovo.csv'
model_HOS = results_path + oversamp + 'rm_bsln/' + 'maxRR/' + 'HOS/' + 'weighted/' + 'C_0.001' + '_decision_ovo.csv'
model_myDesc = results_path + oversamp + 'rm_bsln/' + 'maxRR/' + 'myMorph/' + 'weighted/' + 'C_0.001' + '_decision_ovo.csv'
# Load Predictions!
prob_ovo_RR = np.loadtxt(model_RR)
prob_ovo_wvl = np.loadtxt(model_wvl)
prob_ovo_LBP = np.loadtxt(model_LBP)
prob_ovo_HOS = np.loadtxt(model_HOS)
prob_ovo_MyDescp = np.loadtxt(model_myDesc)
prob_ovo_HBF = np.loadtxt(model_HBF)
predict, prob_ovo_RR_sig = ovo_voting_exp(prob_ovo_RR, 4) #voting_ovo_w(prob_ovo_RR) #voting_ovo_raw(prob_ovo_RR)
predict, prob_ovo_wvl_sig = ovo_voting_exp(prob_ovo_wvl, 4) #voting_ovo_w(prob_ovo_wvl) #voting_ovo_raw(prob_ovo_wvl)
predict, prob_ovo_LBP_sig = ovo_voting_exp(prob_ovo_LBP, 4) #voting_ovo_w(prob_ovo_HOS_myDesc) #voting_ovo_raw(prob_ovo_HOS_myDesc)
predict, prob_ovo_HOS_sig = ovo_voting_exp(prob_ovo_HOS, 4) #voting_ovo_w(prob_ovo_HOS_myDesc) #voting_ovo_raw(prob_ovo_HOS_myDesc)
predict, prob_ovo_MyDescp_sig = ovo_voting_exp(prob_ovo_MyDescp, 4) #voting_ovo_w(prob_ovo_HOS_myDesc) #voting_ovo_raw(prob_ovo_HOS_myDesc)
predict, prob_ovo_HBF_sig = ovo_voting_exp(prob_ovo_HBF, 4)
##########################################################
# Combine the predictions!
##########################################################
# 2
#probs_ensemble = np.stack((prob_ovo_RR_sig, prob_ovo_wvl_sig))
#probs_ensemble = np.stack((prob_ovo_RR_sig, prob_ovo_HOS_sig))
#probs_ensemble = np.stack((prob_ovo_RR_sig, prob_ovo_LBP_sig))
#probs_ensemble = np.stack((prob_ovo_RR_sig, prob_ovo_MyDescp_sig))
#probs_ensemble = np.stack((prob_ovo_wvl_sig, prob_ovo_HOS_sig))
#probs_ensemble = np.stack((prob_ovo_wvl_sig, prob_ovo_LBP_sig))
#probs_ensemble = np.stack((prob_ovo_wvl_sig, prob_ovo_MyDescp_sig))
#probs_ensemble = np.stack((prob_ovo_HOS_sig, prob_ovo_LBP_sig))
#probs_ensemble = np.stack((prob_ovo_HOS_sig, prob_ovo_MyDescp_sig))
#probs_ensemble = np.stack((prob_ovo_LBP_sig, prob_ovo_MyDescp_sig))
# 3
#probs_ensemble = np.stack((prob_ovo_RR_sig, prob_ovo_wvl_sig, prob_ovo_HOS_sig))
#probs_ensemble = np.stack((prob_ovo_RR_sig, prob_ovo_wvl_sig, prob_ovo_LBP_sig))
#probs_ensemble = np.stack((prob_ovo_RR_sig, prob_ovo_wvl_sig, prob_ovo_MyDescp_sig))
#probs_ensemble = np.stack((prob_ovo_RR_sig, prob_ovo_HOS_sig, prob_ovo_LBP_sig))
#probs_ensemble = np.stack((prob_ovo_RR_sig, prob_ovo_HOS_sig, prob_ovo_MyDescp_sig))
#probs_ensemble = np.stack((prob_ovo_RR_sig, prob_ovo_LBP_sig, prob_ovo_MyDescp_sig))
#probs_ensemble = np.stack((prob_ovo_wvl_sig, prob_ovo_HOS_sig, prob_ovo_LBP_sig))
#probs_ensemble = np.stack((prob_ovo_wvl_sig, prob_ovo_HOS_sig, prob_ovo_MyDescp_sig))
#probs_ensemble = np.stack((prob_ovo_wvl_sig, prob_ovo_LBP_sig, prob_ovo_MyDescp_sig))
#probs_ensemble = np.stack((prob_ovo_HOS_sig, prob_ovo_LBP_sig, prob_ovo_MyDescp_sig))
# 4
#probs_ensemble = np.stack((prob_ovo_RR_sig, prob_ovo_wvl_sig, prob_ovo_HOS_sig, prob_ovo_LBP_sig))
#probs_ensemble = np.stack((prob_ovo_RR_sig, prob_ovo_wvl_sig, prob_ovo_HOS_sig, prob_ovo_MyDescp_sig))
#probs_ensemble = np.stack((prob_ovo_RR_sig, prob_ovo_wvl_sig, prob_ovo_LBP_sig, prob_ovo_MyDescp_sig))
#probs_ensemble = np.stack((prob_ovo_RR_sig, prob_ovo_HOS_sig, prob_ovo_LBP_sig, prob_ovo_MyDescp_sig))
#probs_ensemble = np.stack((prob_ovo_wvl_sig, prob_ovo_HOS_sig, prob_ovo_LBP_sig, prob_ovo_MyDescp_sig))
# 5
probs_ensemble = np.stack((prob_ovo_RR_sig, prob_ovo_wvl_sig, prob_ovo_HOS_sig, prob_ovo_LBP_sig, prob_ovo_MyDescp_sig))
n_ensembles, n_instances, n_classes = probs_ensemble.shape
###########################################
# product rule!
predictions_prob_rule = basic_rules(probs_ensemble, 0)
perf_measures = compute_AAMI_performance_measures(predictions_prob_rule.astype(int), eval_labels)
write_AAMI_results( perf_measures, results_path + 'fusion/prod_rule_score_Ijk_' + str(format(perf_measures.Ijk, '.2f')) + '_' + DS + '.txt')
###########################################
# Sum rule!
"""
predictions_sum_rule = basic_rules(probs_ensemble, 1)
perf_measures = compute_AAMI_performance_measures(predictions_sum_rule.astype(int), eval_labels)
write_AAMI_results( perf_measures, results_path + 'fusion/sum_rule_score_Ijk_' + str(format(perf_measures.Ijk, '.2f')) + '_' + DS + '.txt')
"""
# min rule!
"""
predictions_min_rule = basic_rules(probs_ensemble, 2)
perf_measures = compute_AAMI_performance_measures(predictions_min_rule.astype(int), eval_labels)
write_AAMI_results( perf_measures, results_path + 'fusion/min_rule_score_Ijk_' + str(format(perf_measures.Ijk, '.2f')) + '_' + DS + '.txt')
"""
# max rule!
"""
predictions_max_rule = basic_rules(probs_ensemble, 3)
perf_measures = compute_AAMI_performance_measures(predictions_max_rule.astype(int), eval_labels)
write_AAMI_results( perf_measures, results_path + 'fusion/max_rule_score_Ijk_' + str(format(perf_measures.Ijk, '.2f')) + '_' + DS + '.txt')
"""
# Mayority rule / Ranking
"""
predictions_rank_rule = basic_rules(probs_ensemble, 4)
perf_measures = compute_AAMI_performance_measures(predictions_rank_rule.astype(int), eval_labels)
write_AAMI_results( perf_measures, results_path + 'fusion/rank_rule_score_Ijk_' + str(format(perf_measures.Ijk, '.2f')) + '_' + DS + '.txt')
"""
if __name__ == '__main__':
import sys
main()
Datasets
Models
