def train_model(num_steps=num_training_steps):
'''Train the model with minibatches in a tensorflow session'''
with tf.Session(graph=self.graph) as session:
tf.initialize_all_variables().run()
print('Initializing variables...')
for step in range(num_steps):
offset = (step * batch_size + startloc) % (
self.train_Y.shape[0] - batch_size)
batch_data = self.train_X[offset:(offset +
batch_size), :]
batch_labels = self.train_Y[offset:(offset +
batch_size)]
# Data to feed into the placeholder variables in the tensorflow graph
feed_dict = {
tf_train_batch: batch_data,
tf_train_labels: batch_labels,
dropout_keep_prob: dropout_prob
}
_, l, predictions = session.run(
[optimizer, loss, batch_prediction],
feed_dict=feed_dict)
if (step % 200 == 0):
train_preds = session.run(train_prediction,
feed_dict={
tf_train_dataset:
self.train_X,
dropout_keep_prob:
1.0
})
val_preds = session.run(
valid_prediction,
feed_dict={dropout_keep_prob: 1.0})
test_preds = session.run(
test_prediction,
feed_dict={dropout_keep_prob: 1.0})
print('')
print('Batch loss at step %d: %f' % (step, l))
print('Batch training accuracy: %.1f%%' %
accuracy(predictions, batch_labels))
print('Validation accuracy: %.1f%%' %
accuracy(val_preds, self.val_Y))
print('Full train accuracy: %.1f%%' %
accuracy(train_preds, self.train_Y))
#if accuracy(val_preds, self.val_Y)>70:
# break
print('Test accuracy: %.1f%%' %
accuracy(test_preds, self.test_Y))
# for theta in threshold:
n = len(self.val_Y)
matrix = np.zeros((n, 5))
matrix[:, 0] = np.array(self.val_Y).reshape(n) * 2 - 1
matrix[:, 1] = np.array(self.val_noise).reshape(n)
counter = 0
for i in val_preds:
if i > (0.5 + theta / 2):
matrix[counter, 2:5] = [1, 0, 0]
elif i < (0.5 - theta / 2):
matrix[counter, 2:5] = [0, 0, 1]
else:
matrix[counter, 2:5] = [0, 1, 0]
counter += 1
Sensitivity, Specificity, MAcc = DataSet.heart_sound_scoring(
matrix)
# res[0, num], res[1, num], res[2, num] = DataSet.heart_sound_scoring(matrix)
# print('Final Results:')
print('Sensitivity: ', Sensitivity)
print('Specificity: ', Specificity)
print('MAcc: ', MAcc)
counter = 0
n = len(self.test_Y)
matrix = np.zeros((n, 5))
matrix[:, 0] = np.array(self.test_Y).reshape(n) * 2 - 1
matrix[:, 1] = np.array(self.test_noise).reshape(n)
for i in test_preds:
if i > (0.5 + theta / 2):
matrix[counter, 2:5] = [1, 0, 0]
elif i < (0.5 - theta / 2):
matrix[counter, 2:5] = [0, 0, 1]
else:
matrix[counter, 2:5] = [0, 1, 0]
counter += 1
TestSensitivity, TestSpecificity, TestMAcc = DataSet.heart_sound_scoring(
matrix)
# res[0, num], res[1, num], res[2, num] = DataSet.heart_sound_scoring(matrix)
print('Testing Results')
print('Sensitivity: ', TestSensitivity)
print('Specificity: ', TestSpecificity)
print('MAcc: ', TestMAcc)
print('Dropout Probability: %1f' % dropout_prob)
print('Weight Penalty: %1f' % weight_penalty)
print('Number of Steps: %d' % num_training_steps)
if Writetofile:
fd = open(OUTPUT_FILE, 'a')
fd.write('\n %1f, %d, %d, %d, %d,' %
(theta, layer1_depth, layer3_depth,
num_training_steps, batch_size))
fd.write('%1f, %1f,' % (dropout_prob, weight_penalty))
fd.write('%1f%%, %1f%%, %1f%%,' %
(accuracy(train_preds, self.train_Y),
accuracy(val_preds, self.val_Y),
accuracy(test_preds, self.test_Y)))
fd.write(
'%1f%%, %1f%%, %1f%%,' %
(Sensitivity * 100, Specificity * 100, MAcc * 100))
fd.write('%1f%%, %1f%%, %1f%%,' %
(TestSensitivity * 100, TestSpecificity * 100,
TestMAcc * 100))
fd.write('%r,' % balanced)
# fd.write('%r, %r, %r, %1f, %1f, %d, %.1f%%, %.1f%%' % (pooling, Augment, invariance, dropout_prob, weight_penalty, num_training_steps, accuracy(val_preds, self.val_Y), accuracy(train_preds, self.train_Y)))
fd.close()
matrix = np.zeros((predicted_noise_y.size, 5))
matrix[:, 0] = test_sound_y
matrix[:, 1] = test_noise_y
counter = 0
for i in predicted_noise_y:
if i == 1:
if predicted_sound_y[counter] == 1:
matrix[counter, 2:5] = [1, 0, 0]
else:
matrix[counter, 2:5] = [0, 0, 1]
else:
matrix[counter, 2:5] = [0, 1, 0]
counter = counter + 1
Sensitivity, Specificty, MAcc = DataSet.heart_sound_scoring(matrix)
print('Final Results:')
print('Sensitivity: ', Sensitivity)
print('Specificity: ', Specificty)
print('MAcc: ', MAcc)
print('________________________________')
print('Auto on Manual trained Abnormal/Normal')
conf = confusion_matrix(test_sound_y, logreg_sound.predict(test_x))
print(conf)
print('Sensitivity: ', conf[0, 0] / (conf[0, 0] + conf[1, 0]))
print('Specificity: ', conf[1, 1] / (conf[1, 1] + conf[0, 1]))
print('________________________________')
temp = max < theta
temp = temp.astype(int)
counter = 0
for i in temp:
if i == 0:
if label[counter] == 1:
matrix[counter, 2:5] = [1, 0, 0]
else:
matrix[counter, 2:5] = [0, 0, 1]
else:
matrix[counter, 2:5] = [0, 1, 0]
counter += 1
res[0, num], res[1, num], res[2, num] = DataSet.heart_sound_scoring(matrix)
num += 1
# Now on test data
theta = threshold[np.argmax(res[2, :])]
probs = logreg.predict_proba(test_x)
label = logreg.predict(test_x)
max = np.amax(probs, axis=1)
matrix = np.zeros((label.size, 5))
matrix[:, 0] = test_y_sound
matrix[:, 1] = test_y_noise
temp = max < theta
def train_model(num_steps=num_training_steps):
'''Train the model with minibatches in a tensorflow session'''
with tf.Session(graph=self.graph) as session:
tf.initialize_all_variables().run()
print('Initializing variables...')
for step in range(num_steps):
# offset = (step * batch_size+startloc) % (self.train_Y.shape[0] - batch_size)
# batch_data = self.train_X[offset:(offset + batch_size), :]
# batch_labels = self.train_Y[offset:(offset + batch_size)]
#
# # Data to feed into the placeholder variables in the tensorflow graph
# feed_dict = {tf_train_batch: batch_data, tf_train_labels: batch_labels}
rand_index = np.random.choice(len(self.train_Y),
size=batch_size)
batch_data = self.train_X[rand_index, :]
batch_labels = self.train_Y[rand_index]
feed_dict = {
tf_train_batch: batch_data,
tf_train_labels: batch_labels
}
_, l, predictions = session.run(
[train_step, loss, batch_prediction],
feed_dict=feed_dict)
if (step % 200 == 0):
train_preds = session.run(train_prediction,
feed_dict={
tf_train_dataset:
self.train_X,
tf_train_batch:
batch_data,
tf_train_labels:
batch_labels
})
val_preds = session.run(valid_prediction,
feed_dict=feed_dict)
test_preds = session.run(test_prediction,
feed_dict=feed_dict)
print('')
print('Batch loss at step %d: %f' % (step, l))
print('Batch training accuracy: %.1f%%' %
accuracy(predictions, batch_labels))
print('Validation accuracy: %.1f%%' %
accuracy(val_preds, self.val_Y))
print('Full train accuracy: %.1f%%' %
accuracy(train_preds, self.train_Y))
#if accuracy(val_preds, self.val_Y)>70:
# break
print('Test accuracy: %.1f%%' %
accuracy(test_preds, self.test_Y))
n = len(self.val_Y)
matrix = np.zeros((n, 5))
matrix[:, 0] = np.array(self.val_Y).reshape(n)
matrix[:, 1] = np.array(self.val_noise).reshape(n)
counter = 0
for i in val_preds:
if np.sign(i) > theta:
matrix[counter, 2:5] = [1, 0, 0]
elif np.sign(i) < -theta:
matrix[counter, 2:5] = [0, 0, 1]
else:
matrix[counter, 2:5] = [0, 1, 0]
counter += 1
Sensitivity, Specificity, MAcc = DataSet.heart_sound_scoring(
matrix)
# res[0, num], res[1, num], res[2, num] = DataSet.heart_sound_scoring(matrix)
# print('Final Results:')
print('Sensitivity: ', Sensitivity)
print('Specificity: ', Specificity)
print('MAcc: ', MAcc)
print('')
print('Parameters')
print('Number of Steps: %d' % num_training_steps)
if Writetofile:
fd = open(OUTPUT_FILE, 'a')
fd.write('\n %1f, %d, %d, %1f, %1f, %r,' %
(theta, num_training_steps, batch_size, C,
gamma, UsePCA))
fd.write('%1f%%, %1f%%, %1f%%,' %
(accuracy(train_preds, self.train_Y),
accuracy(val_preds, self.val_Y),
accuracy(test_preds, self.test_Y)))
fd.write(
'%1f%%, %1f%%, %1f%%,' %
(Sensitivity * 100, Specificity * 100, MAcc * 100))
#fd.write('%r, %r, %r, %1f, %1f, %d, %.1f%%, %.1f%%' % (pooling, Augment, invariance, dropout_prob, weight_penalty, num_training_steps, accuracy(val_preds, self.val_Y), accuracy(train_preds, self.train_Y)))
fd.close()
