关于深度学习的一点理解

参考大神的文章https://www.zybuluo.com/hanbingtao/note/476663

感觉逐渐入门了深度学习，在这写点现阶段自己关于对神经网络和反向传播算法的理解。简单的神经网络通常有三层：输入层，隐藏层，输出层。如果不考虑复杂的数学公式推导的话，其实没有多少东西，只要弄清每层的输入输出，每层的连接方式，其实并没有多困难，输入输出比较容易弄清楚，输入层的输出就是隐藏层的输入，以此推，很容易理解。连接方式，无非是每一层是全连接还是连接一部分，连接的时候每一个神经元的权重是多少的问题。最难的部分在于根据一定的误差准则去确定权重，这个时候需要复杂的数学公式进行推导，一般是链式求导。在求导的时候会根据一定的算法，比如反向传播算法等，求导完成之后又会进一步去优化，优化的算法又会有很多种，有随机梯度下降算法等（自己还没掌握那么多算法，暂且写自己掌握的）最后得到我们求更新权重的公式，然后根据公式，用python或者c++等等其他任何一种语言去敲出来，大功告成。贴上大神的代码基于python2.7。方便以后进一步学习。

import random
import numpy as np
from activators import SigmoidActivator, IdentityActivator


# 全连接层实现类
class FullConnectedLayer(object):
    def __init__(self, input_size, output_size, 
                 activator):
        '''
        构造函数
        input_size: 本层输入向量的维度
        output_size: 本层输出向量的维度
        activator: 激活函数
        '''
        self.input_size = input_size
        self.output_size = output_size
        self.activator = activator
        # 权重数组W
        self.W = np.random.uniform(-0.1, 0.1,
            (output_size, input_size))
        # 偏置项b
        self.b = np.zeros((output_size, 1))
        # 输出向量
        self.output = np.zeros((output_size, 1))

    def forward(self, input_array):
        '''
        前向计算
        input_array: 输入向量，维度必须等于input_size
        '''
        # 式2
        self.input = input_array
        self.output = self.activator.forward(
            np.dot(self.W, input_array) + self.b)

    def backward(self, delta_array):
        '''
        反向计算W和b的梯度
        delta_array: 从上一层传递过来的误差项
        '''
        # 式8
        self.delta = self.activator.backward(self.input) * np.dot(
            self.W.T, delta_array)
        self.W_grad = np.dot(delta_array, self.input.T)
        self.b_grad = delta_array

    def update(self, learning_rate):
        '''
        使用梯度下降算法更新权重
        '''
        self.W += learning_rate * self.W_grad
        self.b += learning_rate * self.b_grad

    def dump(self):
        print 'W: %s\nb:%s' % (self.W, self.b)


# 神经网络类
class Network(object):
    def __init__(self, layers):
        '''
        构造函数
        '''
        self.layers = []
        for i in range(len(layers) - 1):
            self.layers.append(
                FullConnectedLayer(
                    layers[i], layers[i+1],
                    SigmoidActivator()
                )
            )

    def predict(self, sample):
        '''
        使用神经网络实现预测
        sample: 输入样本
        '''
        output = sample
        for layer in self.layers:
            layer.forward(output)
            output = layer.output
        return output

    def train(self, labels, data_set, rate, epoch):
        '''
        训练函数
        labels: 样本标签
        data_set: 输入样本
        rate: 学习速率
        epoch: 训练轮数
        '''
        for i in range(epoch):
            for d in range(len(data_set)):
                self.train_one_sample(labels[d], 
                    data_set[d], rate)

    def train_one_sample(self, label, sample, rate):
        self.predict(sample)
        self.calc_gradient(label)
        self.update_weight(rate)

    def calc_gradient(self, label):
        delta = self.layers[-1].activator.backward(
            self.layers[-1].output
        ) * (label - self.layers[-1].output)
        for layer in self.layers[::-1]:
            layer.backward(delta)
            delta = layer.delta
        return delta

    def update_weight(self, rate):
        for layer in self.layers:
            layer.update(rate)

    def dump(self):
        for layer in self.layers:
            layer.dump()

    def loss(self, output, label):
        return 0.5 * ((label - output) * (label - output)).sum()

    def gradient_check(self, sample_feature, sample_label):
        '''
        梯度检查
        network: 神经网络对象
        sample_feature: 样本的特征
        sample_label: 样本的标签
        '''

        # 获取网络在当前样本下每个连接的梯度
        self.predict(sample_feature)
        self.calc_gradient(sample_label)

        # 检查梯度
        epsilon = 10e-4
        for fc in self.layers:
            for i in range(fc.W.shape[0]):
                for j in range(fc.W.shape[1]):
                    fc.W[i,j] += epsilon
                    output = self.predict(sample_feature)
                    err1 = self.loss(sample_label, output)
                    fc.W[i,j] -= 2*epsilon
                    output = self.predict(sample_feature)
                    err2 = self.loss(sample_label, output)
                    expect_grad = (err1 - err2) / (2 * epsilon)
                    fc.W[i,j] += epsilon
                    print 'weights(%d,%d): expected - actural %.4e - %.4e' % (
                        i, j, expect_grad, fc.W_grad[i,j])


from bp import train_data_set


def transpose(args):
    return map(
        lambda arg: map(
            lambda line: np.array(line).reshape(len(line), 1)
            , arg)
        , args
    )


class Normalizer(object):
    def __init__(self):
        self.mask = [
            0x1, 0x2, 0x4, 0x8, 0x10, 0x20, 0x40, 0x80
        ]

    def norm(self, number):
        data = map(lambda m: 0.9 if number & m else 0.1, self.mask)
        return np.array(data).reshape(8, 1)

    def denorm(self, vec):
        binary = map(lambda i: 1 if i > 0.5 else 0, vec[:,0])
        for i in range(len(self.mask)):
            binary[i] = binary[i] * self.mask[i]
        return reduce(lambda x,y: x + y, binary)

def train_data_set():
    normalizer = Normalizer()
    data_set = []
    labels = []
    for i in range(0, 256):
        n = normalizer.norm(i)
        data_set.append(n)
        labels.append(n)
    return labels, data_set

def correct_ratio(network):
    normalizer = Normalizer()
    correct = 0.0;
    for i in range(256):
        if normalizer.denorm(network.predict(normalizer.norm(i))) == i:
            correct += 1.0
    print 'correct_ratio: %.2f%%' % (correct / 256 * 100)


def test():
    labels, data_set = transpose(train_data_set())
    net = Network([8, 3, 8])
    rate = 0.5
    mini_batch = 20
    epoch = 10
    for i in range(epoch):
        net.train(labels, data_set, rate, mini_batch)
        print 'after epoch %d loss: %f' % (
            (i + 1),
            net.loss(labels[-1], net.predict(data_set[-1]))
        )
        rate /= 2
    correct_ratio(net)


def gradient_check():
    '''
    梯度检查
    '''
    labels, data_set = transpose(train_data_set())
    net = Network([8, 3, 8])
    net.gradient_check(data_set[0], labels[0])
    return net