def bsoid_nn(feats,
labels,
comp=COMP,
hldout=HLDOUT,
cv_it=CV_IT,
mlp_params=MLP_PARAMS):
"""
Trains MLP classifier
:param feats: 2D array, original feature space, standardized
:param labels: 1D array, GMM output assignments
:param hldout: scalar, test partition ratio for validating MLP performance in GLOBAL_CONFIG
:param cv_it: scalar, iterations for cross-validation in GLOBAL_CONFIG
:param mlp_params: dict, MLP parameters in GLOBAL_CONFIG
:return classifier: obj, MLP classifier
:return scores: 1D array, cross-validated accuracy
"""
if comp == 1:
feats_train, feats_test, labels_train, labels_test = train_test_split(
feats.T, labels.T, test_size=hldout, random_state=23)
logging.info(
'Training feedforward neural network on randomly partitioned {}% of training data...'
.format((1 - hldout) * 100))
classifier = MLPClassifier(**mlp_params)
classifier.fit(feats_train, labels_train)
logging.info(
'Done training feedforward neural network mapping {} features to {} assignments.'
.format(feats_train.shape, labels_train.shape))
logging.info(
'Predicting randomly sampled (non-overlapped) assignments '
'using the remaining {}%...'.format(HLDOUT * 100))
scores = cross_val_score(classifier,
feats_test,
labels_test,
cv=cv_it,
n_jobs=-1)
timestr = time.strftime("_%Y%m%d_%H%M")
if PLOT_TRAINING:
np.set_printoptions(precision=2)
titles_options = [("Non-normalized confusion matrix", None),
("Normalized confusion matrix", 'true')]
titlenames = [("counts"), ("norm")]
j = 0
for title, normalize in titles_options:
disp = plot_confusion_matrix(classifier,
feats_test,
labels_test,
cmap=plt.cm.Blues,
normalize=normalize)
disp.ax_.set_title(title)
print(title)
print(disp.confusion_matrix)
my_file = 'confusion_matrix_{}'.format(titlenames[j])
disp.figure_.savefig(
os.path.join(OUTPUT_PATH,
str.join('', (my_file, timestr, '.svg'))))
j += 1
plt.show()
else:
classifier = []
scores = []
for i in range(len(feats)):
feats_train, feats_test, labels_train, labels_test = train_test_split(
feats[i].T, labels[i].T, test_size=hldout, random_state=23)
logging.info(
'Training feedforward neural network on randomly partitioned {}% of training data...'
.format((1 - hldout) * 100))
clf = MLPClassifier(**mlp_params)
clf.fit(feats_train, labels_train)
classifier.append(clf)
logging.info(
'Done training feedforward neural network mapping {} features to {} assignments.'
.format(feats_train.shape, labels_train.shape))
logging.info(
'Predicting randomly sampled (non-overlapped) assignments '
'using the remaining {}%...'.format(HLDOUT * 100))
sc = cross_val_score(classifier,
feats_test,
labels_test,
cv=cv_it,
n_jobs=-1)
timestr = time.strftime("_%Y%m%d_%H%M")
if PLOT_TRAINING:
np.set_printoptions(precision=2)
titles_options = [("Non-normalized confusion matrix", None),
("Normalized confusion matrix", 'true')]
j = 0
titlenames = [("counts"), ("norm")]
for title, normalize in titles_options:
disp = plot_confusion_matrix(classifier,
feats_test,
labels_test,
cmap=plt.cm.Blues,
normalize=normalize)
disp.ax_.set_title(title)
print(title)
print(disp.confusion_matrix)
my_file = 'confusion_matrix_clf{}_{}'.format(
i, titlenames[j])
disp.figure_.savefig(
os.path.join(OUTPUT_PATH,
str.join('', (my_file, timestr, '.svg'))))
j += 1
plt.show()
logging.info(
'Scored cross-validated feedforward neural network performance.'.
format(feats_train.shape, labels_train.shape))
return classifier, scores
