#!/usr/bin/env python
"""
train_SVM.py
VARPA, University of Coruna
Mondejar Guerra, Victor M.
15 Dec 2017
"""
from evaluation_AAMI import *
from sklearn import svm
from aggregation_voting_strategies import *
# Eval the SVM model and export the results
def eval_crossval_fold(svm_model, features, labels, multi_mode, voting_strategy):
if multi_mode == 'ovo':
decision_ovo = svm_model.decision_function(features)
if voting_strategy == 'ovo_voting':
predict_ovo, counter = ovo_voting(decision_ovo, 4)
elif voting_strategy == 'ovo_voting_both':
predict_ovo, counter = ovo_voting_both(decision_ovo, 4)
elif voting_strategy == 'ovo_voting_exp':
predict_ovo, counter = ovo_voting_exp(decision_ovo, 4)
# svm_model.predict_log_proba svm_model.predict_proba svm_model.predict ...
perf_measures = compute_AAMI_performance_measures(predict_ovo, labels)
return perf_measures
def run_cross_val(features, labels, patient_num_beats, division_mode, k):
print("Runing Cross validation...")
# C_values
# gamma_values
C_values = [0.0001, 0.001, 0.01, 0.1, 1, 10, 50, 100, 200, 1000]
#ijk_scores = np.zeros(len(C_values))
cv_scores = np.zeros(len(C_values))
index_cv = 0
n_classes = 4
for c_svm in C_values:
# for g in g_values...
features_k_fold = [np.array([]) for i in range(k)]
label_k_fold = [np.array([]) for i in range(k)]
################
# PREPARE DATA
################
if division_mode == 'pat_cv':
k = 22
base = 0
for kk in range(k):
features_k_fold[kk] = features[base:base+patient_num_beats[kk]]
label_k_fold[kk] = labels[base:base+patient_num_beats[kk]]
base = patient_num_beats[kk] + 1
# NOTE: 22 k-folds will be very computational cost
# NOTE: division by patient and oversampling couldnt by used!!!!
if division_mode == 'beat_cv':
# NOTE: class sklearn.model_selection.StratifiedKFold(n_splits=3, shuffle=False, random_state=None)[source]
# Stratified K-Folds cross-validator
# Provides train/test indices to split data in train/test sets.
# Thirun_cross_vals cross-validation object is a variation of KFold that returns stratified folds.
# The folds are made by preserving the percentage of samples for each class!!
# Sort features and labels by class ID
features_by_class = {}
for c in range(n_classes):
features_by_class[c] = features[labels == c]
# Then split each instances group in k-folds
# generate the k-folds with the same class ID proportions in each one
instances_class = len(features_by_class[c])
increment = instances_class / k
base = 0
for kk in range(k):
features_k_fold[kk] = np.vstack((features_k_fold[kk], features_by_class[c][base:base + increment])) if features_k_fold[kk].size else features_by_class[c][base:base+increment]
label_k_fold[kk] = np.hstack((label_k_fold[kk], np.zeros(increment) + c)) if label_k_fold[kk].size else np.zeros(increment) + c
base = increment + 1
################
# RUN CROSS VAL
################
for kk in range(k):
# Rotate each iteration one fold for test and the rest for training
# for each k-fold select the train and validation data
val_features = features_k_fold[kk]
val_labels = label_k_fold[kk]
tr_features = np.array([])
tr_labels =np.array([])
for kkk in range(k):
if kkk != kk:
tr_features = np.vstack((tr_features, features_k_fold[kkk])) if tr_features.size else features_k_fold[kkk]# select k fold
tr_labels = np.append(tr_labels, label_k_fold[kkk])
# pipeline = Pipeline([('transformer', scalar), ('estimator', clf)])
# instead of "StandardScaler()"
####################################################33
# Train
multi_mode = 'ovo'
class_weights = {}
for c in range(n_classes):
class_weights.update({c:len(tr_labels) / float(np.count_nonzero(tr_labels == c))})
#class_weight='balanced',
svm_model = svm.SVC(C=c_svm, kernel='rbf', degree=3, gamma='auto',
coef0=0.0, shrinking=True, probability=False, tol=0.001,
cache_size=200, class_weight=class_weights, verbose=False,
max_iter=-1, decision_function_shape=multi_mode, random_state=None)
# Let's Train!
svm_model.fit(tr_features, tr_labels)
#########################################################################
# 4) Test SVM model
# ovo_voting:
# Simply add 1 to the win class
perf_measures = eval_crossval_fold(svm_model, val_features, val_labels, multi_mode, 'ovo_voting_exp')
# TODO evaluar con el propio Ijk?? de esta manera obtendremos el SVM entrenando en maximizar esa medida...
#ijk_scores[index_cv] += perf_measures.Ijk
cv_scores[index_cv] += np.average(perf_measures.F_measure)
# TODO g-mean?
# Zhang et al computes the g-mean. But they computed the g-mean value for each SVM model of the 1 vs 1. NvsS, NvsV, ..., SvsV....
# NOTE we could use the F-measure average from each class??
print("C value (" + str(c_svm) + " Cross val k " + str(kk) + "/" + str(k) + " AVG(F-measure) = " + str( cv_scores[index_cv] / float(kk+1)))
# range k
# beat division
cv_scores[index_cv] /= float(k)# Average this result with the rest of the k-folds
# NOTE: what measure maximize in the cross val????
index_cv += 1
# c_values
return cv_scores, C_values
Datasets
Models
