def weightLearner(self):
super().weightLearner()
vrsavepath = os.path.join('result', 'voterate.txt')
assert (self.cropping == False)
gent = self._read_data
if self.modelstr == 'EEGNet':
_co = {}
elif self.modelstr == 'rawEEGConv' or self.modelstr == 'TSGLEEGNet':
_co = {'TSG': TSG}
else:
_co = {}
voterate_list = []
data = {'x_train': None, 'y_train': None}
for subject in self.subs:
pred_list = []
for path in walk_files(
os.path.join(self.basepath, '{:0>2d}'.format(subject))):
if not self._readed:
for data['x_train'], data['y_train'] in gent(
subject=subject, mode='train'):
self._readed = True
if self.standardizing:
data = self._standardize(data)
model = tf.keras.models.load_model(path, custom_objects=_co)
_pred = model.predict(data['x_train'],
batch_size=self.batch_size,
verbose=self.verbose)
pred_list.append(
np.squeeze(
np.argmax(_pred, axis=1) == np.squeeze(
data['y_train'])))
pred = np.array(pred_list)
lr = LinearRegression(fit_intercept=False)
lr.fit(pred.T, np.squeeze(np.ones_like(data['y_train'])))
self.weight_list[subject - 1] = lr.coef_
voterate_list.append(lr.coef_)
self._readed = False
with open(vrsavepath, 'w+') as f:
sys.stdout = f
print('Bagging Ensemble Vote Rate (Linear Regression)')
for subject, vr in zip(self.subs, voterate_list):
print('Subject {:0>2d}: '.format(subject),
list(map(lambda x: '{:.2f}'.format(x), vr)))
sys.stdout = _console
f.seek(0, 0)
for line in f.readlines():
print(line)
f.close()
def call(self, *args, **kwargs):
gent = self._read_data
if self.modelstr == 'EEGNet':
_co = {}
elif self.modelstr == 'rawEEGConv' or self.modelstr == 'TSGLEEGNet':
_co = {'TSG': TSG}
else:
_co = {}
acc_list = []
kappa_list = []
data = {'x_test': None, 'y_test': None}
for subject in self.subs:
pred_list = []
for path in walk_files(
os.path.join(self.basepath, '{:0>2d}'.format(subject))):
if not self._readed:
for data['x_test'], data['y_test'] in gent(subject=subject,
mode='test'):
self._readed = True
if self.standardizing:
data = self._standardize(data)
model = tf.keras.models.load_model(path, custom_objects=_co)
if self.cropping:
_Pred = []
for cpd in self._cropping_data((data['x_test'], )):
pd = model.predict(cpd, verbose=0)
_Pred.append(
np.argmax(pd, axis=1) == np.squeeze(
data['y_test']))
_Pred = np.array(_Pred)
Pred = []
for i in np.arange(_Pred.shape[1]):
if _Pred[:, i].any():
Pred.append(1)
else:
Pred.append(0)
acc = np.mean(np.array(Pred))
kappa = 0. # None
else:
_pred = model.predict(data['x_test'],
batch_size=self.batch_size,
verbose=self.verbose)
pred_list.append(np.squeeze(np.argmax(_pred, axis=1)))
pred = self.ensemble(pred_list, self.weight_list[subject - 1])
acc_list.append(np.mean(pred == np.squeeze(data['y_test'])))
kappa_list.append(computeKappa(pred, data['y_test']))
self._readed = False
avg_acc = np.mean(acc_list)
avg_kappa = np.mean(kappa_list)
with open(self.resavepath, 'w+') as f:
sys.stdout = f
print(('{0:s} {1:d}-fold Ensemble({2:s}) ' + self.validation_name +
' Accuracy (kappa)').format(self.modelstr, self.kFold,
self.ename))
for i in range(len(self.subs)):
print('Subject {0:0>2d}: {1:.2%} ({2:.4f})'.format(
self.subs[i], acc_list[i], kappa_list[i]))
print('Average : {0:.2%} ({1:.4f})'.format(avg_acc, avg_kappa))
sys.stdout = _console
f.seek(0, 0)
for line in f.readlines():
print(line)
f.close()
acc_list.append(avg_acc)
kappa_list.append(avg_kappa)
return acc_list, kappa_list