梯度下降逻辑回归初版(bgd)

最新推荐文章于 2022-04-13 00:53:20 发布

张一爻

最新推荐文章于 2022-04-13 00:53:20 发布

阅读量212

点赞数

分类专栏： python代码整合

本文链接：https://blog.csdn.net/weixin_43069769/article/details/105655729

版权

python代码整合专栏收录该内容

115 篇文章 17 订阅

订阅专栏

$J(\theta) = - \frac{1}{m} [ \sum_{i=1}^{m} (y^{(i)}\log(yhat_{\theta}(x^{(i)})) + (1-y^{(i)})\log{(1-yhat_{\theta}(x^{(i)}))})] + \lambda \frac{1}{2m} \sum_{j=1}^n \theta_j^2$

$yhat_\theta(X) = g(\theta^T X) = sigmoid(\theta^T X)$

$\lambda \frac{1}{2m} \sum_{j=1}^n \theta_j^2 在神经网络里就是b偏置项$

def h(x,w):
    return np.dot(x,w)

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

def loss(x,y,w,b=0):
    m,n = x.shape
    yhat = sigmoid(h(x,w))
    J = np.dot(np.log(1-yhat).T,-y)-np.dot(np.log(1-yhat).T,1-y)/m+(b/(2*m))*np.sum(np.square(w[1:]))
    return np.isnan(J[0]) == True and np.inf or J[0]

def bgd(x,y,alpha = 0.05,epsilon = 0.00000001,maxloop=100000,b=0):

    m,n =  x.shape
    w = np.zeros((n,1))
    cost = loss(x,y,w,b)
    costs = [cost]
    ws = [w]
    
    for time in range(maxloop):
    
        yhat = sigmoid(np.dot(x,w))
        w = w-alpha*((1/m)*np.dot(x.T,yhat-y))+(b/m)*np.r_[[0], np.squeeze(w[1:])][:,None]
        ws.append(w)
        cost = loss(x,y,w,b)
        costs.append(cost)
        
        lossvar = abs(costs[-1]-costs[-2])
        if lossvar < epsilon:
            return time, costs,ws
        elif np.isinf(lossvar) or np.isnan(lossvar):
            return lossvar
        else:
            continue

## 数据
Xlo = np.linspace(-10,10,10000)[:,None]
Ylo = (np.random.randn(10000)*10)[:,None]
m,n = Xlo.shape
b = 0
x = np.c_[np.ones((m,1)), Xlo]
y = (np.abs(Ylo/10)>1).astype(int)
##调用批量梯度下降
results1 = bgd(x,y)
time, costs,thetas = results1