美文网首页
(四)随机梯度下降

(四)随机梯度下降

作者: 羽天驿 | 来源:发表于2020-04-06 16:11 被阅读0次

一、什么是随机的梯度下降

随机梯度下降.png

优点:梯度下降,计算所有的样本,随机梯度下降,计算一部分,不仅计算速度快,同时,效果好。因为,有了随机性,将异常值,影响降到最低


 # 随机选取了样本中20个样本,进行计算
    # 当数据量比较多时,计算速度更快
    # 随机性,计算的更加准确
    for i in np.random.choice(np.arange(40),size = 20):#计算20个,返回20个偏导数,求平均值
        cost,dw,db = self.loss(X[i,0],y[i,0])
        cost_ += cost/20
        dw_ += dw/20
        db_ += db/20
1-梯度下降更新规则.png
2-sgd随机梯度下降.png

二、原理的推导

  • J(\theta) = \frac{1}{m}(X\theta - y)^T(X\theta - y)

  • J(\theta) = \frac{1}{m}(\theta^TX^T - y^T)(X\theta - y)

  • J(\theta) = \frac{1}{m}(\theta^TX^TX\theta - \theta^TX^Ty - y^TX\theta + y^Ty) 矩阵乘法形式的变换

  • J(\theta) = \frac{1}{m}(\theta^TX^TX\theta - 2y^TX\theta + y^Ty )

  • 对损失函数进行导数求解:

    • J'(\theta) = \frac{1}{m}(2X^TX\theta - 2X^Ty)
    • J'(\theta) = \frac{2}{m}X^T(X\theta - y)
    • \nabla_{\theta}J(\theta) = \frac{2}{m}X^T(X\theta - y)
  • 更新规则有了

    • \theta = \theta - \frac{2\epsilon}{m}X^T(X\theta - y)
  • 随机梯度下降:

  • import numpy as np
    import matplotlib.pyplot as plt
    
    #------------------------------------#
    X = np.linspace(-2,12,40).reshape(-1,1)
    w = np.random.randint(2,12,size = 1)
    b = np.random.randint(-10,10,size = 1)
    y = X*w + b + np.random.randn(40,1)*2.5
    # 将y.reshape(-1)一维的
    y = y.reshape(-1)
    plt.scatter(X,y,color = 'red')
    # 作为训练数据,增加了一列,截距
    X_train = np.concatenate([X,np.ones(shape = (40,1))],axis = 1)
    #--------------------------------------#
    
    def gradient_descent(X,y):
        m = 10# 从40个样本中随机选取10个样本,计算梯度
        theta = np.random.randn(2) # theta中既有斜率,又有截距
        last_theta = theta + 0.1 #记录theta更新后,和上一步的误差
        precision = 1e-4 #精确度
        epsilon = 0.01 #步幅
        count= 0
        while True:
    #         当斜率和截距误差小于万分之一时,退出
            if (np.abs(theta - last_theta) < precision).all():
                break
            if count > 3000:#死循环执行了3000次
                break
    #         更新
            last_theta = theta.copy()
    #     随机梯度下降,梯度是矩阵计算返回的
            index = np.random.choice(np.arange(40),size = m)# index索引,根据随机索引从原数据中取数据
            grad = 2/m*X[index].T.dot(X[index].dot(theta) - y[index])
            theta -= epsilon*grad
            count += 1
        return theta
    w_,b_ = gradient_descent(X_train,y)
    j = lambda x : w_*x + b_
    plt.scatter(X[:,0],y,color = 'red')
    x_test = np.linspace(-2,12,1024) 
    y_ = j(x_test)
    plt.plot(x_test,y_,color = 'green')
    
3-sgd可视化.png

三、代码的实现

(一、随机梯度下降)

导包

import numpy as np
import matplotlib.pyplot as plt
from sklearn.linear_model import LinearRegression

构建数据

X = np.linspace(-2,12,40).reshape(-1,1)
w = np.random.randint(1,9,size = 1)
b = np.random.randint(-5,5,size = 1)
# 增加了噪声
y = w*X + b + np.random.randn(40,1)*2
plt.scatter(X,y)
<matplotlib.collections.PathCollection at 0x1f560183f88>
output_3_1.png

随机梯度下降

抽样,随机抽取10个

总样本是m个,随机抽样m'
m = 40
m' = 10

# 没有使用矩阵,实数
class LinearModel(object):
    def __init__(self):#初始化,随机给定斜率和截距
        self.w = np.random.randn(1)[0]
        self.b = np.random.randn(1)[0]
        
    def model(self,x):# 模型
        return self.w*x + self.b#一元一次线性方程,模型
    
    def loss(self,x,y):#损失,最小二乘法
        cost = (self.model(x) - y)**2 # 损失函数越小越好
#         求解梯度,两个未知数,所以,偏导
        d_w = 2*(self.model(x) - y)*x # 斜率w的偏导
        d_b = 2*(self.model(x) - y)*1 # 截距b的偏导
        return cost,d_w,d_b
    
    def gradient_descent(self,step,d_w,d_b):# 梯度下降
        self.w -= step*d_w # 根据梯度,更新斜率
        self.b -= step*d_b # 根据梯度,更细截距
        
    def fit(self,X,y):#fit 训练模型,将数据交给模型,寻找规律
        precision = 1e-4 # 精确度
        last_w = self.w + 0.01
        last_b = self.b + 0.01
        count = 0
        while True:
            if count > 2000:
                break
            if (np.abs(self.w - last_w) < precision) & (np.abs(self.b - last_b) < precision):
                break
#             更新斜率和截距
            last_w = self.w # 更新之前,先保留,记录
            last_b = self.b
            cost_ = 0
            dw_ = 0
            db_ = 0
            for i in np.random.choice(np.arange(40),size = 20):#计算10个,返回10个偏导数,求平均值
                cost,dw,db = self.loss(X[i,0],y[i,0])
                cost_ += cost/20
                dw_ += dw/20
                db_ += db/20
            self.gradient_descent(0.02,dw_,db_)
            count += 1
            
        print('----------------------',self.w,self.b)
    def predict(self,X):
        return self.model(X)

使用随机梯度下降,可视化

X_test = np.linspace(-2,12,512).reshape(-1,1)

linear = LinearModel()

linear.fit(X,y)

y_ = linear.predict(X_test)

plt.plot(X_test,y_,color = 'green')

plt.scatter(X,y,color = 'red')
---------------------- 3.840141287097452 -3.106354282077542





<matplotlib.collections.PathCollection at 0x1f561a3a908>
output_8_2.png

(二、随机梯度下降矩阵)

import numpy as np
import matplotlib.pyplot as plt
X = np.linspace(-2,12,40).reshape(-1,1)

w = np.random.randint(2,12,size = 1)

b = np.random.randint(-10,10,size = 1)

y = X*w + b + np.random.randn(40,1)*2.5

# 将y.reshape(-1)一维的
y = y.reshape(-1)

plt.scatter(X,y,color = 'red')
<matplotlib.collections.PathCollection at 0x21147ef99c8>
output_1_1.png

用方法,实现梯度下降

m是样本的数量

\nabla_{\theta}J(\theta) = \frac{2}{m}X^T(X\theta - y)

f(x) = b+ w_1x + w_2x^2 + w_3x^3

f(x) = bx^0 + w_1x + w_2x^2 + w_3x^3

f(x) = w_0x^0 + w_1x + w_2x^2 + w_3x^3

对数据X增加了一列,这一列对应着,截距

# 作为训练数据,增加了一列,截距
X_train = np.concatenate([X,np.ones(shape = (40,1))],axis = 1)
X_train
array([[-2.        ,  1.        ],
       [-1.64102564,  1.        ],
       [-1.28205128,  1.        ],
       [-0.92307692,  1.        ],
       [-0.56410256,  1.        ],
       [-0.20512821,  1.        ],
       [ 0.15384615,  1.        ],
       [ 0.51282051,  1.        ],
       [ 0.87179487,  1.        ],
       [ 1.23076923,  1.        ],
       [ 1.58974359,  1.        ],
       [ 1.94871795,  1.        ],
       [ 2.30769231,  1.        ],
       [ 2.66666667,  1.        ],
       [ 3.02564103,  1.        ],
       [ 3.38461538,  1.        ],
       [ 3.74358974,  1.        ],
       [ 4.1025641 ,  1.        ],
       [ 4.46153846,  1.        ],
       [ 4.82051282,  1.        ],
       [ 5.17948718,  1.        ],
       [ 5.53846154,  1.        ],
       [ 5.8974359 ,  1.        ],
       [ 6.25641026,  1.        ],
       [ 6.61538462,  1.        ],
       [ 6.97435897,  1.        ],
       [ 7.33333333,  1.        ],
       [ 7.69230769,  1.        ],
       [ 8.05128205,  1.        ],
       [ 8.41025641,  1.        ],
       [ 8.76923077,  1.        ],
       [ 9.12820513,  1.        ],
       [ 9.48717949,  1.        ],
       [ 9.84615385,  1.        ],
       [10.20512821,  1.        ],
       [10.56410256,  1.        ],
       [10.92307692,  1.        ],
       [11.28205128,  1.        ],
       [11.64102564,  1.        ],
       [12.        ,  1.        ]])

根据矩阵求解的梯度,进行梯度下降

生成系数时,必须考虑形状

def gradient_descent(X,y):
    m = 10# 从40个样本中随机选取10个样本,计算梯度
    theta = np.random.randn(2) # theta中既有斜率,又有截距
    last_theta = theta + 0.1 #记录theta更新后,和上一步的误差
    precision = 1e-4 #精确度
    epsilon = 0.01 #步幅
    count= 0
    while True:
#         当斜率和截距误差小于万分之一时,退出
        if (np.abs(theta - last_theta) < precision).all():
            break
        if count > 3000:#死循环执行了3000次
            break
#         更新
        last_theta = theta.copy()
#     随机梯度下降,梯度是矩阵计算返回的
        index = np.random.choice(np.arange(40),size = m)# index索引,根据随机索引从原数据中取数据
        grad = 2/m*X[index].T.dot(X[index].dot(theta) - y[index])
        theta -= epsilon*grad
        count += 1
    return theta
w_,b_ = gradient_descent(X_train,y)
j = lambda x : w_*x + b_
plt.scatter(X[:,0],y,color = 'red')
x_test = np.linspace(-2,12,1024) 
y_ = j(x_test)
plt.plot(x_test,y_,color = 'green')
[<matplotlib.lines.Line2D at 0x2114b5eb148>]
output_10_1.png

相关文章

网友评论

      本文标题:(四)随机梯度下降

      本文链接:https://www.haomeiwen.com/subject/gfuruhtx.html