CNN的实现与理解

0 写在前

CNN做为经典机器学习的算法之一,是深度学习的基础.为此写了这篇博客,以记录学习中的所得.

1 训练集与训练标签的设置

本文适用于train为[None,X],其中None为样本,X为样本属性大小.labels为热编码后的标签.并使用mnist数据集.数据读取:

import tensorflow as tf
from tensorflow.examples.tutorials.mnist import input_data
mnist = input_data.read_data_sets('MNIST_data',one_hot = True)
# 原始数据占位
data = mnist.test.images
labels = mnist.test.labels

2 激活函数

对于该算法,激活函数是必不可少的,其作用在于,将线性变换进行非线性转化.使得学习更具有泛化能力.本节有relu函数,sigmoid函数,softmax函数,tanh函数

def sigmoid(x):
    return 1.0/(1+np.exp(-x))
def relu(x):
    return (np.abs(x) + x) / 2
def tanh(x):
    return (np.exp(x) - np.exp(-x)) / (np.exp(x) + np.exp(-x))
def softmax(X):
    orig_shape = X.shape
    if len(X.shape)>1:
        X -= np.max(X,axis = 1,keepdims = True)
        a = np.exp(X)
        b = np.sum(np.exp(X),axis=1)
        for i in range(len(b)):
            X[i]=a[i]/b[i]
    else:
        X -=np.max(X,axis=0,keepdims=True)
        X = np.exp(X)/np.sum(np.exp(X),axis=0)
    assert X.shape == orig_shape
    return X 
     

3 前向传播

$$Z_1=W_{1}X+b_1$$ $$H_1=f(Z_1)$$ $$Z_2=W_{2}X+b_2$$ $$H_2=f(Z_2)$$ $$Z_3=W_{3}X+b_3$$ $$H_3=f(Z_3)$$
其参数设置首先确定网络结构,并初始化权值参数,及代码如下:

def weight_bias(layerdims):
    W = {}
    b = {}
    for i in range(1,len(layerdims)):
        W['W'+str(i)] = np.random.randn(layerdims[i-1],layerdims[i])
        b['b'+str(i)] = np.random.randn(layerdims[i],)
    return W,b    
# 前向传播
def forword(data,Weight,bias,layerdims,activation):
    # 非线性输出
    H = {}
    H['H0'] = data
    # 线性输出
    Z = {}
    for i in range(1,len(layerdims)):
        # 
        Z['Z'+str(i)] = np.dot(H['H'+str(i-1)],Weight['W'+str(i)])+bias['b'+str(i)]
        exec("H['H' + str(i)] = " + activation[i-1] + "(Z['Z' + str(i)])")
    return Z,H

损失函数

$$Loss=-\frac{1}{mN}\sum _{j=1}^N\sum _{i=1}^{m}y\log ({\hat{y}}^i+(1-y)\log (1-{\hat{y}}^i))$$
其中m对应一个样本的属性多少.N为样本量.
其实现代码如下:

def loss_function(H,labels):
    lens = len(H)
    n = labels.shape[0]
    m = labels.shape[1]
    H_end = H['H'+str(lens-1)]
    y_ = H_end
    loss = -np.sum(np.sum(labels*np.log(y_)+(1-labels)*np.log(1-y_),axis = 1))/(m*n)    
    return loss

反向传播

为了方便观看，我们把之前整个的一个前向传播的过程复制过来：
$$Z_1=W_{1}X+b_1$$$$H_1=RELU(Z_1)$$$$Z_2=W_2{H}_1+b_2$$$$H_2=RELU(Z_2)$$$$Z_3=W_3{H}_2+b_3$$$$\hat{y}=sigmoid(Z_3)$$

同时，把损失函数也弄过来：
$$J(w,b)=-\frac{1}{mN}\sum _{j=1}^N\sum _{i=1}^{m}y\log ({\hat{y}}^i+(1-y)\log (1-{\hat{y}}^i))$$
注意，下面有的式子，为了直观没有写出来对矩阵求导的转置，写代码的时候需要注意这一点
首先第一件事是对$z_3$进行求导
$$\frac{\partial J}{\partial z_3}=\frac{\partial J}{\partial \hat{y}}\frac{\partial \hat{y}}{\partial z_3}=\hat{y}-y={\delta }_3$$
为什么求这个式子，当然是为了链式相乘了，然后我们开始对参数w和b进行求导：
$$\frac{\partial J}{\partial w_3}=\frac{\partial J}{\partial z_3}\frac{\partial z_3}{\partial w_3}={\delta }_3{H}_2$$
$$\frac{\partial J}{\partial b_3}=\frac{\partial J}{\partial z_3}\frac{\partial z_3}{\partial b_3}={\delta }_3$$
到这里我们完成了对w3和b3这两个参数进行求导，后面的基本上类似，就是运用链式求导的法则一层层往前求即可，我们我们继续向下写。
$$\frac{\partial J}{\partial z_2}=\frac{\partial J}{\partial z_3}\frac{\partial z_3}{\partial H_2}\frac{\partial H_2}{\partial z_2}={\delta }_3{w}_{3}rel{u}^{\prime }(z_2)={\delta }_2$$
$$\frac{\partial J}{\partial w_2}=\frac{\partial J}{\partial z_2}\frac{\partial z_2}{\partial w_2}={\delta }_2{H}_1$$
$$\frac{\partial J}{\partial b_2}=\frac{\partial J}{\partial z_2}\frac{\partial z_2}{\partial b_2}={\delta }_2$$
对于w1和b1也是一样的
$$\frac{\partial J}{\partial z_1}=\frac{\partial J}{\partial z_2}\frac{\partial z_2}{\partial H_1}\frac{\partial H_1}{\partial z_1}={\delta }_2{w}_{2}rel{u}^{\prime }(z_1)={\delta }_1$$
$$\frac{\partial J}{\partial w_1}=\frac{\partial J}{\partial z_1}\frac{\partial z_1}{\partial w_1}={\delta }_{1}x$$
$$\frac{\partial J}{\partial b_1}=\frac{\partial J}{\partial z_1}\frac{\partial z_1}{\partial b_1}={\delta }_1$$

ok 我们完成了所有的反向求导，接下来我们终于可以愉快的写代码了！

# 反向传播
def backward_propagation(X, labels, weight, bias, H, activation):
    m = X.shape[0]
    gradients = {}
    L = len(weight)
    ## 计算导数
    gradients['dZ' + str(L)] = H['H' + str(L)] - labels
    gradients['dW' + str(L)] = 1./m * np.dot( H['H' + str(L-1)].T,gradients['dZ' + str(L)]) 
    gradients['db' + str(L)] = 1./m * np.sum(gradients['dZ' + str(L)].T, axis=1, keepdims = True)
    for l in range(L-1, 0, -1):
        gradients['dH' + str(l)] = np.dot(gradients['dZ'+str(l+1)],weight['W'+str(l+1)].T)
        if activation[l-1] == 'relu':
            gradients['dZ'+str(l)] = np.multiply(gradients['dH'+str(l)], np.int64(H['H'+str(l)] > 0))
        elif activation[l-1] == 'tanh':
            gradients['dZ'+str(l)] = np.multiply(gradients['dH'+str(l)], 1 - np.power(H['H'+str(l)], 2))
        elif activation[l-1] == 'sigmoid':
            gradients['dZ'+str(l)] = np.multiply(gradients['dH'+str(l)], sigmoid_gard(H['H'+str(l)]))
        gradients['dW'+str(l)] = 1./m * np.dot(H['H' + str(l-1)].T,gradients['dZ' + str(l)])
        gradients['db'+str(l)] = 1./m * np.sum(gradients['dZ' + str(l)].T, axis=1, keepdims = True)
    return gradients

本代码先计算最后一层的损失导数,然后递推得到前面每一层的参数导数.最后就是参数更新了

# 参数更新
def updata_parameters(weight, bias, gradients, lr):
    ## 更新参数，lr 为 learning rate，是代表参数的学习率
    ## 太小会使网络收敛很慢，太大可能会使网络在最低点附近徘徊而不会收敛
    for i in range(1, len(weight)+1):
        weight['W'+str(i)] -= lr * gradients['dW'+str(i)]
        bias['b'+str(i)] -= lr * gradients['db'+str(i)][0]

    return weight, bias  

模型训练

def nn_fit(X,Y, Weight, bias, activation, lr, lambd=0.7, num_iterations=5000, print_cost=[True, 100]):
    ## num_iteration 是迭代的次数，print_cost，可以在每几次迭代后打印成本。
    ## 每次迭代都会按「前向传播，计算成本，计算梯度，更新梯度」的顺序执行 
    for i in range(num_iterations):
        Z,H = forword(X,Weight, bias,layerdims,activation)
        cost = loss_function(H,Y)
        grads = backward_propagation(X, Y, Weight, bias, H, activation)
        Weight, b  = updata_parameters(Weight, bias, grads, lr)

        if print_cost[0] and i % print_cost[1] == 0:
            print("Cost after iteration %i: %f" % (i, cost))
    return Weight, bias

模型测试

def cls_predict(X, Weight, bias, activation):
    ## 输出大于 0.5 的视为 1
    Z,H = forword(X,Weight, bias,layerdims,activation)
    prediction = (H['H'+str(len(H)-1)] > 0.5)
    return prediction

具体例子1:

import matplotlib.pyplot as plt
from sklearn.datasets import make_circles
import keras
# In[2]:
# 准备数据
def load_data():
    # 训练样本有 300 个，测试样本有 100 个
    train_X, train_Y = make_circles(n_samples=6600, noise=.02)
    test_X, test_Y = make_circles(n_samples=1200, noise=.02)
    # 可视化数据
    plt.scatter(train_X[:, 0], train_X[:, 1], c=train_Y, s=40, cmap=plt.cm.Spectral);
    train_X = train_X.T
    train_Y = train_Y.reshape((1, train_Y.shape[0]))
    test_X = test_X.T
    test_Y = test_Y.reshape((1, test_Y.shape[0]))
    return train_X, train_Y, test_X, test_Y

train_X, train_Y, test_X, test_Y = load_data()
X = train_X.T
Y = keras.utils.to_categorical(train_Y, 2)[0]
activation = ['relu','relu','sigmoid']
layerdims = [2,18,7,2]
Weight, bias = weight_bias(layerdims)
Weight, bias = nn_fit(X,Y, Weight, bias, activation, lr=0.1, lambd=0.7, num_iterations=5000, print_cost=[True, 100])
X1 = test_X.T
Y1 = keras.utils.to_categorical(test_Y, 2)[0]
prediction = cls_predict(X1, Weight, bias, activation)
accuracy = np.mean((prediction == Y1),dtype=np.float64)
print(accuracy)

具体例子2:

import tensorflow as tf
from tensorflow.examples.tutorials.mnist import input_data
mnist = input_data.read_data_sets('MNIST_data',one_hot = True)
train_X, train_Y, test_X, test_Y = mnist.train.images,mnist.train.labels,mnist.validation.images,mnist.validation.labels
activation = ['sigmoid','sigmoid','sigmoid']
layerdims = [784,256,64,10]
Weight, bias = weight_bias(layerdims)
Weight, bias = nn_fit(train_X, train_Y, Weight, bias, activation, lr=0.05, lambd=0.2, num_iterations=2000, print_cost=[True, 50])
prediction = cls_predict(test_X, Weight, bias, activation)
accuracy = np.mean((prediction== test_Y),dtype=np.float64)
print(accuracy)