4逻辑回归线性不可分

4逻辑回归线性不可分

题目：设想你是工厂的生产主管，你有一些芯片在两次测试中的测试结果。对于这两次测试，你想决定是否芯片要被接受或抛弃
数据集：ex2data2.txt

1. 数据可视化

# 读取文件
data = pd.read_csv("ex2data2.txt", sep=",", names=["test1", "test2", "accepted"])
print(data.head())
# 数据可视化
fig, ax = plt.subplots()
ax.scatter(data[data["accepted"] == 0]["test1"], data[data["accepted"] == 0]["test2"], c= "r", marker="x", label="y=0")
ax.scatter(data[data["accepted"] == 1]["test1"], data[data["accepted"] == 1]["test2"], c= "b", marker="o", label="y=1")
ax.legend()
ax.set(xlabel="test1", ylabel= "test2")
plt.show()

2. 特征映射
   如果样本量多，逻辑回归问题很复杂，而原始特征只有x1,x2可以用多项式创建更多的特征x1、x2、x1x2、x1^2、x22、… X1^nX2n。因为更多的特征进行逻辑回归时，得到的分割线可以是任意高阶函数的形状。

def feature_mapping(x1, x2, power):
    data = {}
    for i in np.arange(power + 1):
        for j in np.arange(i + 1):
            data["F{}{}".format(i-j, j)] = np.power(x1, i-j) * np.power(x2, j)
    return pd.DataFrame(data)   # 转成dataframe结构


x1 = data["test1"]
x2 = data["test2"]
data2 = feature_mapping(x1, x2, 6)
print(data2.head())

3. 正则化损失函数
   
   不惩罚第一项，j从1开始

X = data2.values # (118,28)
y = data.iloc[:, -1].values
y = y.reshape(len(y), 1) # (118,1)


def sigmoid(z):
    return 1 / (1 + np.exp(-z))


def cost_function(X, y, theta, lamda):
    A = sigmoid(X @ theta)
    first = y * np.log(A)
    second = (1-y) * np.log(1 - A)
    reg = np.sum(np.power(theta[1:], 2)) * (lamda / (2 * len(X)))
    return -np.sum(first + second) / len(X) + reg


theta = np.zeros((28, 1))
lamda = 1
cost_init = cost_function(X, y, theta, lamda)
print(cost_init) # 0.693

4. 梯度下降函数
   

def gradient_descent(X, y, theta, alpha, iters, lamda):
    costs = []
    for i in range(iters):
        reg = theta[1:] * (lamda / len(X))  # 现在是27行一列
        reg = np.insert(reg, 0, values=0, axis=0) # 变为28行1列
        theta = theta - (X.T @ (sigmoid(X @ theta) - y)) * alpha /len(X) - reg
        cost = cost_function(X, y, theta, lamda)
        costs.append(cost)
        # 抽取几个cost查看，避免太多
        if i % 10000 == 0:
            print(costs)
    return costs, theta


alpha = 0.001
iters = 200000
lamda = 0.001
costs， theta_final = gradient_descent(X, y, theta, alpha, iters, lamda)
print(costs) #0.469

5. 准确率

def predict(X, theta):
    prob = sigmoid(X @ theta)
    return [1 if x >= 0.5 else 0 for x in prob]


y_ = np.array(predict(X,theta_final))
y_pre  = y_.reshape(len(y_), 1)
acc = np.mean(y_pre == y)
print(acc) # 0.83

6. 决策边界

x = np.linspace(-1.2, 1.2, 200)
xx, yy = np.meshgrid(x, x)
z = feature_mapping(xx.ravel(), yy.ravel(), 6).values
zz = z @ theta_final
zz = zz.reshape(xx.shape)

fig, ax = plt.subplots()
ax.scatter(data[data["accepted"] == 0]["test1"], data[data["accepted"] == 0]["test2"], c= "r", marker="x", label="y=0")
ax.scatter(data[data["accepted"] == 1]["test1"], data[data["accepted"] == 1]["test2"], c= "b", marker="o", label="y=1")
ax.legend()
ax.set(xlabel="test1", ylabel= "test2")
plt.contour(xx, yy, zz, 0)
plt.show()

改变lamda可改变图形效果