input_ = input_[np.newaxis,:]
target = np.array([target])[np.newaxis,:]
net.activation(input_, accumulate=True)
errors_valid[i] = net.output - target
i += 1
mean_valid, std_valid = (errors_valid.mean(), errors_valid.std())
# test
p_anomaly_test = np.zeros(shape=len(X_test))
predictions = np.zeros(shape=len(X_test))
errors_test = np.zeros(shape=len(Y_test))
threshold = utils.gaussian_pdf(mean_valid-2*std_valid, mean_valid, std_valid)
i = 0
for (input_, target) in zip(X_test, Y_test):
# format input and prediction
input_ = input_[np.newaxis,:]
target = np.array([target])[np.newaxis,:]
net.activation(input_, accumulate=True)
prediction = net.output
predictions[i] = prediction
errors_test[i] = scistats.norm.ppf(prediction-target, mean_valid, std_valid)
anomalies = np.argwhere(errors_test < threshold)
net.activation(input_, accumulate=True)
# execute the backpropagation with the input that has been memorized previously
net.backpropagation(target=target,
loss='L2',
optimizer='sgd',
l_rate=1e-2,
update=True)
i += 1
# test
p_anomaly_test = np.zeros(shape=len(X_test))
predictions = np.zeros(shape=len(X_test))
errors_test = np.zeros(shape=len(Y_test))
threshold = utils.gaussian_pdf(mean_valid - 2. * std_valid, mean_valid,
std_valid)
i = 0
for (input_, target) in zip(X_test, Y_test):
# format input and prediction
input_ = input_[np.newaxis, :]
target = np.array([target])[np.newaxis, :]
net.activation(input_, accumulate=True)
prediction = net.output
predictions[i] = prediction
errors_test[i] = utils.gaussian_pdf(prediction - target, mean_valid,
std_valid)
anomalies = np.argwhere(errors_test < threshold)