12.24 深度学习神经网络代码实现

简单的深度学习其实就是多层的神经网络。接下使用代码实现简单的深度神经网络。

#输入的layer_num是一个list[]
def initialize_parameters(layer_num): 
        l = len(layer_num)  #计算有几层,，把输入也看作一层
        w = {}  #创造两个dict 
        b = {}
        for i in range(1,l):
            w[str(i)] = np.random.randn(layer_num[i],layer_num[i-1])*0.01 
            b[str(i)] = np.zeros((layer_num[i],1)) #这里需要两个括号，不然不识别
        return w,b
def linear_activation(A,w,b):
        Z = np.dot(w,A) + b
        A = sigmoid(Z)
        return Z,A
def forward_pass(X,w,b,layer_num):
        cachez = {}
        cachea = {}
        cachea[str(0)] = X
        l = len(layer_num) 
        for i in range(1,l):
            cachez[str(i)],cachea[str(i)] = linear_activation(cachea[str(i-1)],w[str(i)],b[str(i)])          
            A = cachea[str(i)]
        return A,cachez,cachea
def calculate_cost(A,Y):
        m = A.shape[0] #计算Y一维的长度
        cost = (1.0/m)*((-np.dot(Y,np.log(A).T))-np.dot(1-Y,np.log(1-A).T))
        cost = np.squeeze(cost) #去除秩为1的
        return cost
def back_propogation(Y,A,Z,a,cachea):
        l = len(Z)
        dZ={}
        dZ[str(l-1)] = A - Y
        m = A.shape[0]
        for i in range(l-1,0,-1):
            dw = (1.0/m)*(np.dot(dZ[str(i)],cachea[str(i-1)].T))
            db = (1.0/m)*(np.sum(dZ[str(i)],axis=1,keepdims=True))
            w[str(i)] -= a*dw
            b[str(i)] -= a*db
            dA = np.dot(w[str(i)].T,dZ[str(i)])
            dZ[str(i)] = np.multiply(dA,sigmoid_derivative(Z[str(i-1)]))
            return w,b  
def train(X,Y,layer_num,learn_rate=0.2,repeat=10):
    np.random.seed(5) #给定随机数
    a = 0.2
    w,b = initialize_parameters(layer_num)
    for i in range(1,repeat): 
            A,Z,cachea = forward_pass(X,w,b,layer_num)
            cost = calculate_cost(A,Y)
            back_propogation(Y,A,Z,a,cachea)

我们可以实验一下，训练一个。

X = np.array([[0, 0], [0, 1], [1, 0], [1, 1]])
Y = np.array([0, 1])
layer_num=[4,3,6,1]
train(X,Y,layer_num)

之后可以def一个类来专门输入test数据。
这里可以发现矩阵的行列数字会成为编程中的关键因素。所以shape语句会很重要，但是在练习中没有都使用。

Z.shape == (W.shape[0], A.shape[1])