b站刘二大人----梯度下降法

梯度：

如何更新权重（迭代）

什么是非凸函数？

其存在局部最优点。梯度下降法是无法避免局部最优的，但是为什么常用梯度下降法呢，因为在深度学习中局部最优点是比较少，而深度学习更需要解决的是鞍点问题。

 什么是鞍点？

可以认为梯度为0的点.如下所示。当梯度下降法遇到鞍点的时候，它是无法迭代的

 更新过程

其中，α是学习率

代码如下：

import matplotlib.pyplot as plt
 
# prepare the training set
x_data = [1.0, 2.0, 3.0]
y_data = [2.0, 4.0, 6.0]
 
# initial guess of weight 
w = 1.0
 
# define the model linear model y = w*x
def forward(x):
    return x*w
 
#define the cost function MSE 
def cost(xs, ys):
    cost = 0
    for x, y in zip(xs,ys):
        y_pred = forward(x)
        cost += (y_pred - y)**2
    return cost / len(xs)
 
# define the gradient function  gd
def gradient(xs,ys):
    grad = 0
    for x, y in zip(xs,ys):
        grad += 2*x*(x*w - y)
    return grad / len(xs)
 
epoch_list = []
cost_list = []
print('predict (before training)', 4, forward(4))
for epoch in range(100):
    cost_val = cost(x_data, y_data)
    grad_val = gradient(x_data, y_data)
    w-= 0.01 * grad_val  # 0.01 learning rate
    print('epoch:', epoch, 'w=', w, 'loss=', cost_val)
    epoch_list.append(epoch)
    cost_list.append(cost_val)
 
print('predict (after training)', 4, forward(4))
plt.plot(epoch_list,cost_list)
plt.ylabel('cost')
plt.xlabel('epoch')
plt.show() 

结果：

predict (before training) 4 4.0
epoch: 0 w= 1.0933333333333333 loss= 4.666666666666667
epoch: 1 w= 1.1779555555555554 loss= 3.8362074074074086
epoch: 2 w= 1.2546797037037036 loss= 3.1535329869958857
epoch: 3 w= 1.3242429313580246 loss= 2.592344272332262
epoch: 4 w= 1.3873135910979424 loss= 2.1310222071581117
epoch: 5 w= 1.4444976559288012 loss= 1.7517949663820642
epoch: 6 w= 1.4963445413754464 loss= 1.440053319920117
epoch: 7 w= 1.5433523841804047 loss= 1.1837878313441108
epoch: 8 w= 1.5859728283235668 loss= 0.9731262101573632
epoch: 9 w= 1.6246153643467005 loss= 0.7999529948031382
epoch: 10 w= 1.659651263674342 loss= 0.6575969151946154
epoch: 11 w= 1.6914171457314033 loss= 0.5405738908195378
epoch: 12 w= 1.7202182121298057 loss= 0.44437576375991855
epoch: 13 w= 1.7463311789976905 loss= 0.365296627844598
epoch: 14 w= 1.7700069356245727 loss= 0.3002900634939416
epoch: 15 w= 1.7914729549662791 loss= 0.2468517784170642
epoch: 16 w= 1.8109354791694263 loss= 0.2029231330489788
epoch: 17 w= 1.8285815011136133 loss= 0.16681183417217407
epoch: 18 w= 1.8445805610096762 loss= 0.1371267415488235
epoch: 19 w= 1.8590863753154396 loss= 0.11272427607497944
epoch: 20 w= 1.872238313619332 loss= 0.09266436490145864
epoch: 21 w= 1.8841627376815275 loss= 0.07617422636521683
epoch: 22 w= 1.8949742154979183 loss= 0.06261859959338009
epoch: 23 w= 1.904776622051446 loss= 0.051475271914629306
epoch: 24 w= 1.9136641373266443 loss= 0.04231496130368814
epoch: 25 w= 1.9217221511761575 loss= 0.03478477885657844
epoch: 26 w= 1.9290280837330496 loss= 0.02859463421027894
epoch: 27 w= 1.9356521292512983 loss= 0.023506060193480772
epoch: 28 w= 1.9416579305211772 loss= 0.01932302619282764
epoch: 29 w= 1.9471031903392007 loss= 0.015884386331668398
epoch: 30 w= 1.952040225907542 loss= 0.01305767153735723
epoch: 31 w= 1.9565164714895047 loss= 0.010733986344664803
epoch: 32 w= 1.9605749341504843 loss= 0.008823813841374291
epoch: 33 w= 1.9642546069631057 loss= 0.007253567147113681
epoch: 34 w= 1.9675908436465492 loss= 0.005962754575689583
epoch: 35 w= 1.970615698239538 loss= 0.004901649272531298
epoch: 36 w= 1.9733582330705144 loss= 0.004029373553099482
epoch: 37 w= 1.975844797983933 loss= 0.0033123241439168096
epoch: 38 w= 1.9780992835054327 loss= 0.0027228776607060357
epoch: 39 w= 1.980143350378259 loss= 0.002238326453885249
epoch: 40 w= 1.9819966376762883 loss= 0.001840003826269386
epoch: 41 w= 1.983676951493168 loss= 0.0015125649231412608
epoch: 42 w= 1.9852004360204722 loss= 0.0012433955919298103
epoch: 43 w= 1.9865817286585614 loss= 0.0010221264385926248
epoch: 44 w= 1.987834100650429 loss= 0.0008402333603648631
epoch: 45 w= 1.9889695845897222 loss= 0.0006907091659248264
epoch: 46 w= 1.9899990900280147 loss= 0.0005677936325753796
epoch: 47 w= 1.9909325082920666 loss= 0.0004667516012495216
epoch: 48 w= 1.9917788075181404 loss= 0.000383690560742734
epoch: 49 w= 1.9925461188164473 loss= 0.00031541069384432885
epoch: 50 w= 1.9932418143935788 loss= 0.0002592816085930997
epoch: 51 w= 1.9938725783835114 loss= 0.0002131410058905752
epoch: 52 w= 1.994444471067717 loss= 0.00017521137977565514
epoch: 53 w= 1.9949629871013967 loss= 0.0001440315413480261
epoch: 54 w= 1.9954331083052663 loss= 0.0001184003283899171
epoch: 55 w= 1.9958593515301082 loss= 9.733033217332803e-05
epoch: 56 w= 1.9962458120539648 loss= 8.000985883901657e-05
epoch: 57 w= 1.9965962029289281 loss= 6.57716599593935e-05
epoch: 58 w= 1.9969138906555615 loss= 5.406722767150764e-05
epoch: 59 w= 1.997201927527709 loss= 4.444566413387458e-05
epoch: 60 w= 1.9974630809584561 loss= 3.65363112808981e-05
epoch: 61 w= 1.9976998600690001 loss= 3.0034471708953996e-05
epoch: 62 w= 1.9979145397958935 loss= 2.4689670610172655e-05
epoch: 63 w= 1.9981091827482769 loss= 2.0296006560253656e-05
epoch: 64 w= 1.9982856590251044 loss= 1.6684219437262796e-05
epoch: 65 w= 1.9984456641827613 loss= 1.3715169898293847e-05
epoch: 66 w= 1.9985907355257035 loss= 1.1274479219506377e-05
epoch: 67 w= 1.9987222668766378 loss= 9.268123006398985e-06
epoch: 68 w= 1.9988415219681517 loss= 7.61880902783969e-06
epoch: 69 w= 1.9989496465844576 loss= 6.262999634617916e-06
epoch: 70 w= 1.9990476795699081 loss= 5.1484640551938914e-06
epoch: 71 w= 1.9991365628100501 loss= 4.232266273994499e-06
epoch: 72 w= 1.999217150281112 loss= 3.479110977946351e-06
epoch: 73 w= 1.999290216254875 loss= 2.859983851026929e-06
epoch: 74 w= 1.9993564627377531 loss= 2.3510338359374262e-06
epoch: 75 w= 1.9994165262155628 loss= 1.932654303533636e-06
epoch: 76 w= 1.999470983768777 loss= 1.5887277332523938e-06
epoch: 77 w= 1.9995203586170245 loss= 1.3060048068548734e-06
epoch: 78 w= 1.9995651251461022 loss= 1.0735939958924364e-06
epoch: 79 w= 1.9996057134657994 loss= 8.825419799121559e-07
epoch: 80 w= 1.9996425135423248 loss= 7.254887315754342e-07
epoch: 81 w= 1.999675878945041 loss= 5.963839812987369e-07
epoch: 82 w= 1.999706130243504 loss= 4.902541385825727e-07
epoch: 83 w= 1.9997335580874436 loss= 4.0301069098738336e-07
epoch: 84 w= 1.9997584259992822 loss= 3.312926995781724e-07
epoch: 85 w= 1.9997809729060159 loss= 2.723373231729343e-07
epoch: 86 w= 1.9998014154347876 loss= 2.2387338352920307e-07
epoch: 87 w= 1.9998199499942075 loss= 1.8403387118941732e-07
epoch: 88 w= 1.9998367546614149 loss= 1.5128402140063082e-07
epoch: 89 w= 1.9998519908930161 loss= 1.2436218932547864e-07
epoch: 90 w= 1.9998658050763347 loss= 1.0223124683409346e-07
epoch: 91 w= 1.9998783299358769 loss= 8.403862850836479e-08
epoch: 92 w= 1.9998896858085284 loss= 6.908348768398496e-08
epoch: 93 w= 1.9998999817997325 loss= 5.678969725349543e-08
epoch: 94 w= 1.9999093168317574 loss= 4.66836551287917e-08
epoch: 95 w= 1.9999177805941268 loss= 3.8376039345125727e-08
epoch: 96 w= 1.9999254544053418 loss= 3.154680994333735e-08
epoch: 97 w= 1.9999324119941766 loss= 2.593287985380858e-08
epoch: 98 w= 1.9999387202080534 loss= 2.131797981222471e-08
epoch: 99 w= 1.9999444396553017 loss= 1.752432687141379e-08
predict (after training) 4 7.999777758621207

 上面只是理想情况，而实际情况中会存在噪音，通常的结果可能是这样：

 有的时候震荡很大的时候。我们会做一个指数加权均值的方法使得更加平滑，如下所示：

 如果你的cost图像出现如下情况，说明训练发散（失败），此时的最低点并不是你需要的那个点，造成这个有很多原因，其中包括学习率太大。

 随机梯度下降

 

之前我们使用的是每个样本的损失函数来对w求导。而随机梯度下降是某一个样本的损失函数对w求导。我们的数据是有噪声的，所以这里就是随机噪声，当我们陷入鞍点的时候，可能会跨越鞍点，达到最优。

普通梯度下降效率高但是性能不高。随机梯度下降虽然性能高，能找到最优点，但由于它的每一个w之间都是互相联系的，不能并行，所以时间复杂度高。为了折中，所以产生了batch，即小批量随机梯度下降。

随机梯度下降法和梯度下降法的主要区别在于：

1、损失函数由cost()更改为loss()。cost是计算所有训练数据的损失，loss是计算一个训练数据的损失。对应于源代码则是少了两个for循环。

2、梯度函数gradient()由计算所有训练数据的梯度更改为计算一个训练数据的梯度。
3、本算法中的随机梯度主要是指，每次拿一个训练数据来训练，然后更新梯度参数。本算法中梯度总共更新100(epoch)x3 = 300次。梯度下降法中梯度总共更新100(epoch)次。————————————————
版权声明：本文为CSDN博主「错错莫」的原创文章，遵循CC 4.0 BY-SA版权协议，转载请附上原文出处链接及本声明。
原文链接：https://blog.csdn.net/bit452/article/details/109637108

import matplotlib.pyplot as plt
 
x_data = [1.0, 2.0, 3.0]
y_data = [2.0, 4.0, 6.0]
 
w = 1.0
 
def forward(x):
    return x*w
 
# calculate loss function
def loss(x, y):
    y_pred = forward(x)
    return (y_pred - y)**2

------------------
#之前的损失函数
def cost(xs, ys):
    cost = 0
    for x, y in zip(xs,ys):
        y_pred = forward(x)
        cost += (y_pred - y)**2
    return cost / len(xs)
------------------

# define the gradient function  sgd
def gradient(x, y):
    return 2*x*(x*w - y)

------------------
#之前的梯度下降
#def gradient(xs,ys):
    grad = 0
    for x, y in zip(xs,ys):
        grad += 2*x*(x*w - y)
    return grad / len(xs)
------------------

epoch_list = []
loss_list = []
print('predict (before training)', 4, forward(4))
for epoch in range(100):
    for x,y in zip(x_data, y_data):
        grad = gradient(x,y)
        w = w - 0.01*grad    # update weight by every grad of sample of training set
        print("\tgrad:", x, y,grad)
        l = loss(x,y)
    print("progress:",epoch,"w=",w,"loss=",l)
    epoch_list.append(epoch)
    loss_list.append(l)
 
print('predict (after training)', 4, forward(4))
plt.plot(epoch_list,loss_list)
plt.ylabel('loss')
plt.xlabel('epoch')
plt.show()