exercise10_binary_logistic_regression.py

# Sascha Spors, Professorship Signal Theory and Digital Signal Processing,
# Institute of Communications Engineering (INT), Faculty of Computer Science
# and Electrical Engineering (IEF), University of Rostock, Germany
#
# Data Driven Audio Signal Processing - A Tutorial with Computational Examples
# Feel free to contact lecturer frank.schultz@uni-rostock.de
#
# Exercise 10: Binary logistic regression model with just one layer
# Training using gradient descent and forward/backward propagation
# following the derivations and coding conventions from the brilliant
# course https://www.coursera.org/learn/neural-networks-deep-learning
# cf. especially week 2

import numpy as np

# rng = np.random.RandomState(1)  # for debug
rng = np.random.RandomState()


def my_sigmoid(x):
    return 1 / (1 + np.exp(-x))


def cost(y, a):
    # vectorized loss
    L = -(y * np.log(a) + (1 - y) * np.log(1 - a))
    # cost function as avg of all L entries
    J = np.mean(L)
    return J


def evaluate(y_true, y_pred):
    # actually not nice, # since we generally overwrite y_pred
    y_pred[y_pred < 0.5], y_pred[y_pred >= 0.5] = 0, 1
    # which might get dangerous if we need original data outside the function
    # therefore we should call: evaluate(np.copy(), np.copy())

    # inverted logic to be consistent with the TF confusion matrix
    # of labels starting with 0:
    # label positive == 0
    # label negative == 1
    # confusion matrix (row = actual label, column = predicted label):
    # [TP    FN] = [0,0    0,1<-this 1 is negative label and hene false]
    # [FP    TN] = [1,0    1,1]
    TP = np.sum(np.logical_and(np.logical_not(y_true), np.logical_not(y_pred)))
    TN = np.sum(np.logical_and(y_true, y_pred))
    FN = np.sum(np.logical_xor(y_true[y_true == 0], y_pred[y_true == 0]))
    FP = np.sum(np.logical_xor(y_true[y_true == 1], y_pred[y_true == 1]))
    cm = np.array([[TP, FN], [FP, TN]])
    n_cm = cm / np.sum(np.sum(cm))

    # sensitivity, recall, hit rate, or true positive rate (TPR)
    recall = TP / (TP + FN)
    # specificity, selectivity or true negative rate (TNR)
    TN / (TN + FP)
    # precision or positive predictive value (PPV)
    precision = TP / (TP + FP)
    # negative predictive value (NPV)
    TN / (TN + FN)
    # accuracy (ACC)
    accuracy = (TP + TN) / (TP + TN + FP + FN)
    # balanced F-score, F1 score
    F1_score = 2 / (1 / precision + 1 / recall)  # harmonic mean
    return cm, n_cm, precision, recall, F1_score, accuracy


def create_data(nx, m, mean_offset):
    # std=1,  all features have same +mean
    X1 = rng.randn(nx, m // 2) + mean_offset
    # vary the variance [0...1] of the features
    for i in range(nx):
        X1[i, :] *= rng.rand(1)
    # class 1 code
    Y1 = np.ones((1, m // 2))

    X2 = rng.randn(nx, m // 2) - mean_offset
    for i in range(nx):
        X2[i, :] *= rng.rand(1)
        Y2 = np.zeros((1, m // 2))

    X = np.concatenate((X1, X2), axis=1)
    Y = np.concatenate((Y1, Y2), axis=1)
    return X, Y


# gradient descent hyper parameters
# set up to see how things evolve
# in practice do automated hyper parameter search
step_size = 0.5
steps = 500

# we create some training data
mean_offset = 0.75  # apply +-mean_offset to randn data to simulate two classes
m = 100000  # training data examples, even int!
m_test = m // 10  # test data examples, even int!
nx = 5  # number of features, increasing
X, Y = create_data(nx, m, mean_offset)

# we init weights and bias
w = (rng.rand(nx, 1) - 1 / 2) * 2
b = (rng.rand(1, 1) - 1 / 2) * 2

# we do gradient descent
for step in range(steps):
    # forward propagation = calc actual prediction, i.e. the model output
    # using the current weights and bias:
    Z = np.dot(w.T, X) + b  # forward step 1 = inner product + bias
    A = my_sigmoid(Z)  # forward step 2 = activation function
    print(
        "epoch", step, "/", steps, ", cost function on train data", cost(Y, A)
    )

    # backward propagation, start at the output from model and subsequently
    # move back to the model input
    # vectorized
    da = -Y / A + (1 - Y) / (1 - A)  # step1 -> dL/da
    dz = da * A * (1 - A)  # step 2 -> (dL/da) * da/dz
    dw = np.dot(X, dz.T) / m  # step 3 -> dL/dw = (dL/da * da/dz) * dz/dw
    db = np.mean(dz)  # dL/dw = dL/da * da/dz * dz/db

    # gradient descent update rule
    w = w - step_size * dw
    b = b - step_size * db

# we trained the model and hope for useful weights and bias
print("w", w, "\nb", b)

# check model prediction on training data
print("\n\nmetrics on train data:")
A = my_sigmoid(np.dot(w.T, X) + b)
print("cost function:", cost(Y, A))
A[A < 0.5], A[A >= 0.5] = 0, 1
cm, n_cm, precision, recall, F1_score, accuracy = evaluate(Y, A)
print("accuray", accuracy)
print("precision", precision)
print("recall", recall)
print("F1_score", F1_score * 100, "%")
print("confusion matrix\n[TP FN]\n[FP TN] =\n", n_cm * 100, "%")


# we check model performance on ! unseen ! test data
# therefore we create some test data with same
# PDF characteristics as training data
X, Y = create_data(nx, m_test, mean_offset)

# check model prediction on test data
print("\n\nmetrics on test data:")
A = my_sigmoid(np.dot(w.T, X) + b)
print("cost function:", cost(Y, A))
A[A < 0.5], A[A >= 0.5] = 0, 1
cm, n_cm, precision, recall, F1_score, accuracy = evaluate(Y, A)
print("accuray", accuracy)
print("precision", precision)
print("recall", recall)
print("F1_score", F1_score * 100, "%")
print("confusion matrix\n[TP FN]\n[FP TN] =\n", n_cm * 100, "%")