吴恩达L2W1-作业1

code：

from planar_utils import plot_decision_boundary
import numpy as np
import h5py
import matplotlib.pyplot as plt
from dnn_utils_v2 import relu_backward
from create_data import load_data
class dnn:
    def __init__(self,layer_dims,init_mod='random') -> None:
        '''layer_dimes:各层的神经元数量
            init_mod:初始化方法:zeros,random,he三种初始化'''
        self.WL={}
        self.bL={}
        self.L=len(layer_dims)-1
        # 初始化参数
        if init_mod=='random':
            for i in range(1,self.L+1):
                
                self.WL['W'+str(i)]=np.random.randn(layer_dims[i], layer_dims[i-1]) 
                self.bL['b'+str(i)]=np.zeros((layer_dims[i],1))
        elif init_mod=='zeros':
            for i in range(1,self.L+1):
                self.WL['W'+str(i)]=np.zeros((layer_dims[i], layer_dims[i-1]))
                self.bL['b'+str(i)]=np.zeros((layer_dims[i],1))
        else:
            for i in range(1,self.L+1):
                self.WL['W'+str(i)]=np.random.randn(layer_dims[i], layer_dims[i-1])*np.sqrt(2./layer_dims[i-1])
                self.bL['b'+str(i)]=np.zeros((layer_dims[i],1))

        self.XL={}
        self.AL={}
        self.ZL={}
        self.dZ={}
        self.dW={}
        self.db={}
        self.dA={}
    def input_data(self,X,Y,learning_rate):
        self.m=X.shape[1]
        self.AL["A0"]=X
        self.Y=Y
        self.learning_rate=learning_rate
    def set_data(self,X):
        self.m=X.shape[1]
        self.AL["A0"]=X
    #下面是前向传播模块,前向传播过程中需要记录Z,A 
    def linear_activation_forward(self,i,activation):
        
        '''实现一层的正向传播'''
        # self.ZL[Zi]和self.AL[Ai]记录第i层的数据
        # 存储了计算的Z ,A,W,b,对象自带的有
        self.ZL['Z'+str(i)]=np.dot(self.WL['W'+str(i)],self.AL['A'+str(i-1)])+self.bL['b'+str(i)]
        if activation=="sigmoid":
            self.AL['A'+str(i)]=1/(1+np.exp(-self.ZL['Z'+str(i)]))
        elif activation=="relu":
            self.AL['A'+str(i)]=np.maximum(0,self.ZL['Z'+str(i)])    
        
    def L_model_forward(self):
        # 前L-1层使用relu函数激活,最后一层使用sigmoid函数激活
        for i in range(1,self.L):
            self.linear_activation_forward(i,"relu")
        self.linear_activation_forward(self.L,"sigmoid")
        # 确定最后的输出是否是二分类所需要的输出
        assert(self.AL['A'+str(self.L)].shape==(1,self.m))

    # 下面是计算损失函数
    def computer_cost(self):
        return np.squeeze(-1/self.m*np.sum(self.Y*np.log(self.AL['A'+str(self.L)]+1e-5)+(1-self.Y)*np.log(1-self.AL['A'+str(self.L)]+1e-5)))

    # 下面是反向传播模块
    def linear_backward(self,i):
        '''根据dz[L]计算dw[L],db[L],dA[L-1]'''
        self.dW['dW'+str(i)]=1/self.m*np.dot(self.dZ['dZ'+str(i)],self.AL["A"+str(i-1)].T)
        self.db['db'+str(i)]=1/self.m*np.sum(self.dZ['dZ'+str(i)],axis=1,keepdims=True)
        self.dA['dA'+str(i-1)]=np.dot(self.WL['W'+str(i)].T,self.dZ['dZ'+str(i)])
    def L_model_backforward(self):
        #  先计算最后一层
        # 计算dA
        self.dA['dA'+str(self.L)]=-(np.divide(self.Y,self.AL['A'+str(self.L)]+1e-10)-np.divide(1-self.Y,1-self.AL['A'+str(self.L)]+1e-10))
        # 计算dz
        s=1/(1+np.exp(-self.ZL['Z'+str(self.L)]))
        self.dZ['dZ'+str(self.L)]=self.dA['dA'+str(self.L)]*s*(1-s)
        # 计算dw,db,dA[L-1]
        self.linear_backward(self.L)
        for i in reversed(range(self.L)):
            if i==0:
                break
            else:
                self.dZ['dZ'+str(i)]=relu_backward(self.dA['dA'+str(i)],self.ZL['Z'+str(i)])
                # 计算当前i层的dw,db,和dA[L-1]
                self.linear_backward(i)   

    # 更新参数
    def update_wb(self):
        for i in range(1,self.L+1):
            self.WL['W'+str(i)]=self.WL['W'+str(i)]-self.learning_rate*self.dW['dW'+str(i)]
            self.bL['b'+str(i)]=self.bL['b'+str(i)]-self.learning_rate*self.db['db'+str(i)]

    def train(self,iterations):
        costs=[]
        for i in range(iterations):
            # 前向传播
            self.L_model_forward()
            # 计算损失
            cost=self.computer_cost()
            # 后向传播
            self.L_model_backforward()
            # 更新参数
            self.update_wb()
            if i%100==0:
                costs.append(cost)
                print("第"+str(i)+"次迭代cost:"+str(cost))
            
        return costs
    def predict(self,X):
        self.set_data(X)
        self.L_model_forward()
        return self.AL['A'+str(self.L)]>=0.5


#加载数据
x, y = load_data()

train_num=int(0.8*x.shape[1])
# 划分数据集
train_x=x[:,:train_num]
test_x=x[:,train_num:]
train_y=y[:,:train_num]
test_y=y[:,train_num:]

# 4层的神经网络
my_dnn=dnn([2, 10,  5, 1],'random')

# 设置数据
my_dnn.input_data(train_x,train_y,0.01)

# 训练
cost=my_dnn.train(50000)


# 预测

y_predict=my_dnn.predict(x)
print(y_predict.shape)
print(np.sum(np.abs(y_predict-y)))
plt.subplot(1,2,1)
plot_decision_boundary(lambda x: my_dnn.predict(x.T), x, y)
plt.subplot(1,2,2)
plt.plot(cost)
plt.show()

随机初始化：

zeros初始化：

he初始化：

总结：
零初始化：未能打破对称性
随机初始化：打破了对称性，但权重太大也不行
he初始化：挺好的