def fit(self, X, y, val_X, val_y):
checkpoint = ModelCheckpoint(filepath='m4.hdf5',
verbose=1,
save_best_only=True,
save_weights_only=True,
monitor='val_acc',
mode='max')
callbacks = [checkpoint]
self.clf = Sequential()
self.clf.add(Dense(256, input_dim=X.shape[1],
init='uniform')) #, use_bias=True))
self.clf.add(advanced_activations.ELU(alpha=1.0))
self.clf.add(Dropout(0.6))
self.clf.add(Dense(128, init='uniform'))
self.clf.add(advanced_activations.ELU(alpha=1.0))
self.clf.add(Dropout(0.6))
self.clf.add(Dense(64, use_bias=True, init='uniform'))
self.clf.add(advanced_activations.ELU(alpha=1.0))
self.clf.add(Dropout(0.6))
#model.add(Dense(4, activation='sigmoid', use_bias=True))
#model.add(Dropout(0.5))
self.clf.add(Dense(1, activation='sigmoid', init='uniform'))
adam = Adam(lr=0.001, decay=1e-6, clipvalue=0.5)
self.clf.compile(loss='binary_crossentropy',
optimizer=adam,
metrics=['accuracy'])
return self.clf.fit(X,
y,
epochs=50,
batch_size=2800,
validation_data=(val_X, val_y),
verbose=1,
callbacks=callbacks)
def ResolvedSyn(x=None, type='sigmoid'):
if type == 'sigmoid':
resolved = Sigmoid(x)
elif type == 'tanh':
resolved = Activation('tanh')(x)
elif type == 'relu':
resolved = ReLU(x)
elif type == 'leakyrelu':
resolved = LeakyReLU(x)
elif type == 'elu':
resolved = advanced_activations.ELU()(x)
else:
raise TypeError
print("activated:", resolved)
return resolved
print('x_train shape:', x_train.shape)
print(x_train.shape[0], 'train samples')
print(x_test.shape[0], 'test samples')
y_test_np = np.array(y_test)
y_test_np = y_test_np.flatten()
# Convert class vectors to binary class matrices.
y_train = keras.utils.to_categorical(y_train, num_classes)
y_test = keras.utils.to_categorical(y_test, num_classes)
adam = keras.optimizers.Adam(lr=0.0001,
beta_1=0.9,
beta_2=0.999,
epsilon=1e-08,
decay=0.00001)
ELU_advanced = aact.ELU(alpha=1.1)
regularizer = keras.regularizers.l1(0.001)
tensorboard_caller = keras.callbacks.TensorBoard(log_dir='./logs/Result',
histogram_freq=0,
write_graph=True,
write_images=False)
check_point = keras.callbacks.ModelCheckpoint(filepath='./weights_Result.hdf5',
monitor='val_loss',
verbose=1,
save_best_only=True,
save_weights_only=False,
mode='auto',
period=1)
padding = K.cast( K.not_equal(x[1],0.0), dtype=K.floatx())
return x[0]*padding
if __name__ == "__main__":
X_train, X_val, y_train, y_val = open_data('./UCR_TS_Archive_2015')
input_dim = X_train.shape[-1] # 13
timesteps = X_train.shape[1] # 3
batch_size = 32
# Create graph structure.
input_placeholder = Input(shape=(None, input_dim))
# Encoder.
masked_input = Masking(mask_value=0.0, input_shape=(None,input_dim), name = 'masking_layer')(input_placeholder)
encoded = LSTM(60, return_sequences=True, dropout = 0.2,unit_forget_bias=True)(masked_input)
encoded = advanced_activations.ELU(alpha=.5)(encoded)
encoded = LSTM(60,return_sequences=True, dropout = 0.2, unit_forget_bias=True)(encoded)
encoded = advanced_activations.ELU(alpha=.5)(encoded)
encoded = LSTM(60,dropout = 0.2, unit_forget_bias=True)(encoded)
encoded = advanced_activations.ELU(alpha=.5)(encoded)
encoded = Dense(5)(encoded)
encoded = advanced_activations.ELU(alpha=.5)(encoded)
encoded_final = BatchNormalization(name='embedding')(encoded)
# Decoder.
decoded = Lambda(repeat)([masked_input,encoded_final])
decoded = LSTM(60, return_sequences=True, dropout = 0.2, unit_forget_bias=True)(decoded)
decoded = advanced_activations.ELU(alpha=.5)(decoded)
decoded = LSTM(60, return_sequences=True, dropout = 0.2, unit_forget_bias=True)(decoded)
decoded = advanced_activations.ELU(alpha=.5)(decoded)
decoded = LSTM(input_dim, return_sequences=True, dropout = 0.2, unit_forget_bias=True)(decoded)
def main(config):
# read and preprocess of data
df = pd.read_csv(config.get('GENERAL', 'inputfile'))
df['Amount_max_fraud'] = 1
df.loc[df.Amount <= 2125.87, 'Amount_max_fraud'] = 0
df = df.drop([
'V28', 'V27', 'V26', 'V25', 'V24', 'V23', 'V22', 'V20', 'V15', 'V13',
'V8'
],
axis=1)
df['V1_'] = df.V1.map(lambda x: 1 if x < -3 else 0)
df['V2_'] = df.V2.map(lambda x: 1 if x > 2.5 else 0)
df['V3_'] = df.V3.map(lambda x: 1 if x < -4 else 0)
df['V4_'] = df.V4.map(lambda x: 1 if x > 2.5 else 0)
df['V5_'] = df.V5.map(lambda x: 1 if x < -4.5 else 0)
df['V6_'] = df.V6.map(lambda x: 1 if x < -2.5 else 0)
df['V7_'] = df.V7.map(lambda x: 1 if x < -3 else 0)
df['V9_'] = df.V9.map(lambda x: 1 if x < -2 else 0)
df['V10_'] = df.V10.map(lambda x: 1 if x < -2.5 else 0)
df['V11_'] = df.V11.map(lambda x: 1 if x > 2 else 0)
df['V12_'] = df.V12.map(lambda x: 1 if x < -2 else 0)
df['V14_'] = df.V14.map(lambda x: 1 if x < -2.5 else 0)
df['V16_'] = df.V16.map(lambda x: 1 if x < -2 else 0)
df['V17_'] = df.V17.map(lambda x: 1 if x < -2 else 0)
df['V18_'] = df.V18.map(lambda x: 1 if x < -2 else 0)
df['V19_'] = df.V19.map(lambda x: 1 if x > 1.5 else 0)
df['V21_'] = df.V21.map(lambda x: 1 if x > 0.6 else 0)
df = df.rename(columns={'Class': 'Fraud'})
y = df.Fraud
y = y.values
y = list(y)
print('df', df.shape)
df = df.drop(['Fraud'], axis=1)
X = df.values
X = list(X)
#splitdata
train_X, train_y, val_X, val_y\
= Utils.valicut(X, y, float(config.get('VALIDATE', 'ratio')))
Utils.verbose_print('Standardizing.')
train_X = np.array(train_X)
val_X = np.array(val_X)
index = []
for i in range(19):
index.append(i)
a = train_X[:, index]
scaler = StandardScaler().fit(a)
a = scaler.transform(a)
b = val_X[:, index]
b = scaler.transform(b)
train_X[:, index] = a
val_X[:, index] = b
# sample
Utils.verbose_print('Sampling, method = {0}.'.format(
config.get('SAMPLER', 'method')))
smp = sampler.Smp(config)
train_X, train_y = smp.fit_sample(train_X, train_y)
Utils.verbose_print('data size: {0}.'.format(len(train_y)))
print(train_X.shape)
print(val_X.shape)
#building model
train_y = np.array(train_y)
val_y = np.array(val_y)
early = EarlyStopping('val_acc', patience=0)
checkpoint = ModelCheckpoint(filepath='m3.hdf5'.format(i),
verbose=1,
save_best_only=True,
save_weights_only=True,
monitor='val_acc',
mode='max')
callbacks = [checkpoint]
model = Sequential()
model.add(Dense(256, input_dim=train_X.shape[1],
init='uniform')) #, use_bias=True))
model.add(advanced_activations.ELU(alpha=1.0))
model.add(Dropout(0.6))
model.add(Dense(128, init='uniform'))
model.add(advanced_activations.ELU(alpha=1.0))
model.add(Dropout(0.6))
model.add(Dense(64, use_bias=True, init='uniform'))
model.add(advanced_activations.ELU(alpha=1.0))
model.add(Dropout(0.6))
#model.add(Dense(4, activation='sigmoid', use_bias=True))
#model.add(Dropout(0.5))
model.add(Dense(1, activation='sigmoid', init='uniform'))
adam = Adam(lr=0.001, decay=1e-6, clipvalue=0.5)
model.compile(loss='binary_crossentropy',
optimizer=adam,
metrics=['accuracy'])
model.fit(train_X,
train_y,
epochs=50,
batch_size=2800,
validation_data=(val_X, val_y),
verbose=1,
callbacks=callbacks)
model.load_weights("m3.hdf5")
# validate
Utils.verbose_print('Validating.')
#result_y = clf.predict(val_X)
result_y = model.predict(val_X)
print(result_y.shape)
y = []
for i in range(len(result_y)):
y.append(round(result_y[i][0]))
print(y)
result_y = y
correction = Utils.correction(val_y, result_y)
truth_table = Utils.truth_table(val_y, result_y)
print('Correction:{0}'.format(correction))
Utils.verbose_print(Utils.print_truth_table(truth_table))
f1_score, Precision, Recall = Utils.f1_score(truth_table)
Utils.verbose_print('F1 Score:{0}'.format(f1_score))
Utils.verbose_print('Precision:{0}'.format(Precision))
Utils.verbose_print('Recall:{0}'.format(Recall))
* Optimizer function is 'Adam'
'''
'''
* define numbers for deep learning
'''
FEATURE_NUM = 36
CLASSES = 1
HIDDEN1_SIZE = 100
HIDDEN2_SIZE = 50
MAX_RANGE = 1000
model = Sequential()
model.add(Dense(HIDDEN1_SIZE, input_dim=FEATURE_NUM, init='uniform'))
model.add(advanced_activations.ELU(alpha=1.0))
model.add(Dropout(0.6))
model.add(Dense(HIDDEN2_SIZE, init='uniform'))
model.add(advanced_activations.ELU(alpha=1.0))
model.add(Dropout(0.6))
model.add(Dense(CLASSES, init='uniform', activation='relu'))
'''
* tensorboard and checkpoints saver callbacks
* Keras Tensorboard graph is not prettier than original Tensorflow graph, but much easier to use.
'''
checkpointer = ModelCheckpoint(filepath="/tmp/weights.hdf5",
verbose=1,
save_best_only=True)
tensorboard = TensorBoard(log_dir='./logs',
histogram_freq=0,
def EERACN(shape, NOL):
"""
Network proposed in the paper by Bing Xu et. al.
https://arxiv.org/pdf/1505.00853.pdf
"""
model = Sequential()
model.add(
Conv2D(192, (5, 5),
input_shape=shape,
kernel_initializer='he_normal',
bias_initializer='zeros'))
model.add(BatchNormalization(axis=3))
# model.add(advanced_activations.LeakyReLU(alpha=0.31))
model.add(advanced_activations.ELU())
model.add(
Conv2D(160, (1, 1),
kernel_initializer='he_normal',
bias_initializer='zeros'))
model.add(BatchNormalization(axis=3))
# model.add(advanced_activations.LeakyReLU(alpha=0.31))
model.add(advanced_activations.ELU())
model.add(
Conv2D(96, (1, 1),
kernel_initializer='he_normal',
bias_initializer='zeros'))
model.add(BatchNormalization(axis=3))
#model.add(advanced_activations.LeakyReLU(alpha=0.31))
model.add(advanced_activations.ELU())
model.add(MaxPooling2D(pool_size=(3, 3), strides=(2, 2)))
model.add(Dropout(0.5))
model.add(
Conv2D(192, (5, 5),
kernel_initializer='he_normal',
bias_initializer='zeros'))
model.add(BatchNormalization(axis=3))
# model.add(advanced_activations.LeakyReLU(alpha=0.31))
model.add(advanced_activations.ELU())
model.add(
Conv2D(192, (1, 1),
kernel_initializer='he_normal',
bias_initializer='zeros'))
model.add(BatchNormalization(axis=3))
# model.add(advanced_activations.LeakyReLU(alpha=0.31))
model.add(advanced_activations.ELU())
model.add(
Conv2D(192, (1, 1),
kernel_initializer='he_normal',
bias_initializer='zeros'))
model.add(BatchNormalization(axis=3))
# model.add(advanced_activations.LeakyReLU(alpha=0.31))
model.add(advanced_activations.ELU())
model.add(MaxPooling2D(pool_size=(3, 3), strides=(2, 2)))
model.add(Dropout(0.5))
model.add(
Conv2D(192, (3, 3),
kernel_initializer='he_normal',
bias_initializer='zeros'))
model.add(BatchNormalization(axis=3))
# model.add(advanced_activations.LeakyReLU(alpha=0.31))
model.add(advanced_activations.ELU())
model.add(
Conv2D(192, (1, 1),
kernel_initializer='he_normal',
bias_initializer='zeros'))
model.add(BatchNormalization(axis=3))
# model.add(advanced_activations.LeakyReLU(alpha=0.31))
model.add(advanced_activations.ELU())
model.add(
Conv2D(10, (1, 1),
kernel_initializer='he_normal',
bias_initializer='zeros'))
model.add(BatchNormalization(axis=3))
# model.add(advanced_activations.LeakyReLU(alpha=0.31))
model.add(advanced_activations.ELU())
model.add(
AveragePooling2D(pool_size=(8, 8), strides=(1, 1), padding='same'))
model.add(Flatten())
model.add(Dense(NOL, activation='softmax'))
return model
def Neural_net(x_train_res, y_train_res, method, feature_count):
global test
global test_labels
global x_val
global y_val
model = Sequential()
model.add(Dense(50, kernel_initializer='uniform', input_dim=feature_count))
model.add(advanced_activations.ELU(alpha=1.0))
model.add(Dropout(0.8))
model.add(Dense(30, kernel_initializer='uniform'))
model.add(advanced_activations.ELU(alpha=1.0))
model.add(Dropout(0.9))
model.add(Dense(15, kernel_initializer='uniform', activation="relu"))
# model.add(Dense(15, kernel_initializer='uniform'))
# model.add(advanced_activations.ELU(alpha=1.0))
# model.add(Dense(10, init='uniform', activation=''))
model.add(Dense(1, kernel_initializer='uniform', activation='sigmoid'))
# Compile model
model.compile(loss='binary_crossentropy',
optimizer='adam',
metrics=['sparse_categorical_accuracy'])
model.fit(x_train_res,
y_train_res,
epochs=10,
batch_size=1000,
validation_data=(x_val, y_val))
# output = model.predict_classes(x_val)
# output = [i[0] for i in output]
# output = numpy.array(output)
# print(output[y_val == 1])
# print(sum(output[y_val == 1]))
# print(sum(output))
# print(sum(y_val))
# print(recall_score(y_val, output))
# print(precision_score(y_val, output))
# scores = model.evaluate(train_data, train_labels)
# print("\n%s: %.2f%%" % (model.metrics_names[1], scores[1] * 100))
print("complete")
##Computing false and true positive rates
output = model.predict_classes(test)
output = [i[0] for i in output]
output = numpy.array(output)
print(output[test_labels == 1])
print(sum(output[test_labels == 1]))
print(sum(output))
print(sum(test_labels))
print(recall_score(test_labels, output))
print(precision_score(test_labels, output))
fpr, tpr, _ = roc_curve(output, test_labels, drop_intermediate=False)
if sum(numpy.isnan(tpr)) == 0:
roc_acc = roc_auc_score(output, test_labels)
roc_acc = "Accuracy: " + str(round(roc_acc, 3))
else:
roc_acc = "Accuracy: 0%"
plt.figure()
##Adding the ROC
plt.plot(fpr, tpr, color='red', lw=2, label='ROC curve')
##Random FPR and TPR
plt.plot([0, 1], [0, 1], color='blue', lw=2, linestyle='--')
##Title and label
plt.xlabel('FPR')
plt.ylabel('TPR')
plt.title('Neural Network - (' + method + ")" + str(roc_acc))
# plt.show()
plt.savefig(fname=method + ".png", pad_inches=0.2)
