《PyTorch深度学习实践》学习笔记：梯度下降

最新推荐文章于 2023-01-01 23:42:18 发布

lizhuangabby

最新推荐文章于 2023-01-01 23:42:18 发布

阅读量312

点赞数 1

分类专栏： pytorch 文章标签：深度学习 pytorch 学习

本文链接：https://blog.csdn.net/lizhuangabby/article/details/125536834

版权

pytorch 专栏收录该内容

13 篇文章 4 订阅

订阅专栏

提示：文章写完后，目录可以自动生成，如何生成可参考右边的帮助文档

文章目录

一、优化问题
二、代码练习
总结

一、优化问题

求代价函数最小值
示例：pandas 是基于NumPy 的一种工具，该工具是为了解决数据分析任务而创建的。
沿着负梯度的方向寻找最小值：
在这里插入图片描述
寻找过程类似贪心算法：

对于非凸函数，存在局部最优：

甚至有时会陷入鞍点：

梯度计算：

全局梯度下降->随机梯度下降

全局梯度下降：

随机梯度下降：

随机梯度下降主要是为了解决在机器学习过程中在遇到“鞍点”（即总体所有点的梯度和为0，导致w=w-0.01*0,w不会改变）而导致不能继续进行的问题。可以采用随机梯度下降，即随机的取一组（x,y)的梯度，作为梯度下降的依据，而不用总体所有点的梯度和，作为梯度下降的依据。
在这里插入图片描述
随机梯度下降法在神经网络中被证明是有效的。它的特点在于效率较低(时间复杂度较高)，学习性能较好。

二、代码练习

梯度下降

import numpy as np
import matplotlib.pyplot as plt

x_data = [1.0,2.0,3.0]
y_data = [2.0,4.0,6.0]

w = 1.0
# 前馈计算
def forward(x):
    return x * w

#代价函数
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)

# 梯度计算
def gradident(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 = gradident(x_data,y_data)
    w -= 0.01 * grad_val
    print("Epoch:",epoch,'w=',w,'loss=',cost_val)
    epoch_list.append(epoch+1)
    cost_list.append(cost_val)
print("Predict (after training)",4,forward(4))


# 画图
plt.plot(epoch_list,cost_list)
plt.xlabel("epoch")
plt.ylabel("cost")
plt.show()

运行结果：
在这里插入图片描述
运算输出结果：
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

随机梯度下降：

import numpy as np
import matplotlib.pyplot as plt

x_data = [1.0,2.0,3.0]
y_data = [2.0,4.0,6.0]

w = 1.0
# 前馈计算
def forward(x):
    return x * w

#代价函数
def loss(x,y):
    y_pred = forward(x)
    return (y_pred-y)**2

# 梯度计算
def gradident(x,y):
    grad = 2 * x * (x * w -y)
    return grad

epoch_list = []
cost_list = []
print("Predict (before training)",4,forward(4))
for epoch in range(100):
    for x,y in zip(x_data,y_data):
        grad = gradident(x,y)
        w -= 0.01 * grad
        print("\tgrad:",x,y,grad)
        l = loss(x,y)
    print("progress:", epoch, "w=", w, "loss=", l)
    epoch_list.append(epoch+1)
    cost_list.append(l)

print("Predict (after training)",4,forward(4))


# 画图
plt.plot(epoch_list,cost_list)
plt.xlabel("epoch")
plt.ylabel("loss")
plt.show()

运行结果：
在这里插入图片描述
运行输出结果：
Predict (before training) 4 4.0
grad: 1.0 2.0 -2.0
grad: 2.0 4.0 -7.84
grad: 3.0 6.0 -16.2288
progress: 0 w= 1.260688 loss= 4.919240100095999
grad: 1.0 2.0 -1.478624
grad: 2.0 4.0 -5.796206079999999
grad: 3.0 6.0 -11.998146585599997
progress: 1 w= 1.453417766656 loss= 2.688769240265834
grad: 1.0 2.0 -1.093164466688
grad: 2.0 4.0 -4.285204709416961
grad: 3.0 6.0 -8.87037374849311
…
…
progress: 97 w= 1.9999999999998603 loss= 1.757455879087579e-25
grad: 1.0 2.0 -2.793321129956894e-13
grad: 2.0 4.0 -1.0942358130705543e-12
grad: 3.0 6.0 -2.2648549702353193e-12
progress: 98 w= 1.9999999999998967 loss= 9.608404711682446e-26
grad: 1.0 2.0 -2.0650148258027912e-13
grad: 2.0 4.0 -8.100187187665142e-13
grad: 3.0 6.0 -1.6786572132332367e-12
progress: 99 w= 1.9999999999999236 loss= 5.250973729513143e-26
Predict (after training) 4 7.9999999999996945

总结

随机梯度下降法相对于梯度下降法的改进在于，不再取所有x、y值计算出的损失函数的平均值，而是随机取一组x和y计算损失函数。此方法的作用在于：数据中大多含有随机噪声，则利用这种随机噪声避免鞍点时无法继续迭代的情况。

但是此代码并没有体现随机性，利用zip函数把每一对x和y值都计算了一次（若是输入数据很多则计算量极大），此外，计算出的l和w每次都利用x和y循环中最后一次的结果，即固定取x=3,y=6，与随机性没有任何关系，同时造成前两对x和y值的计算完全无效，属于浪费资源。这里的数据的数量比较小，没有随机取，数量大的时候还是需要严格的随机取才能体现出随机梯度下降的优点。

此外，因为梯度下降和随机梯度下降对于时间和性能各有各的优势，所以进行了折中，提出了批量的梯度下降也就是Mini-Batch梯度下降。