本文使用通俗的语言和形象的图示,介绍了随机梯度下降算法和它的三种经典变体,并提供了完整的实现代码。
GitHub 链接:https://github.com/ManuelGonzalezRivero/3dbabove
代价函数的多种优化方法
目标函数是衡量预测值和实际值的相似程度的指标。通常,我们希望得到使代价尽可能小的参数集,而这意味着你的算法性能不错。函数的最小可能代价被称为最小值。有时一个代价函数可以有多个局部极小值。幸运的是,在参数空间的维数非常高的情况下,阻碍目标函数充分优化的局部最小值并不经常出现,因为这意味着对象函数相对于每个参数在训练过程的早期都是凹的。但这并非常态,通常我们得到的是许多鞍点,而不是真正的最小值。
找到生成最小值的一组参数的算法被称为优化算法。我们发现随着算法复杂度的增加,则算法倾向于更高效地逼近最小值。我们将在这篇文章中讨论以下算法:
- 随机梯度下降法
- 动量算法
- RMSProp
- Adam 算法
随机梯度下降法
我的「Logistic 回归深入浅出」的文章里介绍了一个随机梯度下降如何运作的例子。如果你查阅随机梯度下降法的资料(SGD),通常会遇到如下的等式:
资料上会说,θ是你试图找到最小化 J 的参数,这里的 J 称为目标函数。最后,我们将学习率记为α。通常要反复应用上述等式,直到达到你所需的代价值。
这是什么意思?想一想,假如你坐在一座山顶上的雪橇上,望着另一座山丘。如果你滑下山丘,你会自然地往下移动,直到你最终停在山脚。如果第一座小山足够陡峭,你可能会开始滑上另一座山的一侧。从这个比喻中你可以想到:
学习率越高意味着摩擦力越小,因此雪橇会像在冰上一样沿着山坡下滑。低的学习率意味着摩擦力高,所以雪橇会像在地毯上一样,难以滑下。我们如何用上面的方程来模拟这种效果?
随机梯度下降法:
- 初始化参数(θ,学习率)
- 计算每个θ处的梯度
- 更新参数
- 重复步骤 2 和 3,直到代价值稳定
让我们用一个简单的例子来看看它是如何运作的!
在这里我们看到一个目标函数和它的导数(梯度):
我们可以用下面的代码生成函数和梯度值/30 的图:
-
import numpy as np -
-
def minimaFunction(theta): -
return np.cos(3*np.pi*theta)/theta -
-
def minimaFunctionDerivative(theta): -
const1 = 3*np.pi -
const2 = const1*theta -
return -(const1*np.sin(const2)/theta)-np.cos(const2)/theta**2 -
-
theta = np.arange(.1,2.1,.01) -
Jtheta = minimaFunction(theta) -
dJtheta = minimaFunctionDerivative(theta) -
-
plt.plot(theta,Jtheta,label = r'$J(\theta)$') -
plt.plot(theta,dJtheta/30,label = r'$dJ(\theta)/30$') -
plt.legend() -
axes = plt.gca() -
#axes.set_ylim([-10,10]) -
-
plt.ylabel(r'$J(\theta),dJ(\theta)/30$') -
plt.xlabel(r'$\theta$') -
plt.title(r'$J(\theta),dJ(\theta)/30 $ vs $\theta$') -
plt.show() -
上图中有两个细节值得注意。首先,注意这个代价函数有几个极小值(大约在 0.25、1.0 和 1.7 附近取得)。其次,注意在最小值处的导数在零附近的曲线走向。这个点就是我们所需要的新参。
我们可以在下面的代码中看到上面四个步骤的实现。它还会生成一个视频,显示每个步骤的θ和梯度的值。
-
import numpy as np -
import matplotlib.pyplot as plt -
import matplotlib.animation as animation -
-
def optimize(iterations, oF, dOF,params,learningRate): -
""" -
computes the optimal value of params for a given objective function and its derivative -
Arguments: -
- iteratoins - the number of iterations required to optimize the objective function -
- oF - the objective function -
- dOF - the derivative function of the objective function -
- params - the parameters of the function to optimize -
- learningRate - the learning rate -
Return: -
- oParams - the list of optimized parameters at each step of iteration -
""" -
oParams = [params] -
#The iteration loop -
for i in range(iterations): -
# Compute the derivative of the parameters -
dParams = dOF(params) -
# Compute the update -
params = params-learningRate*dParams -
-
# app end the new parameters -
oParams.append(params) -
-
return np.array(oParams) -
-
def minimaFunction(theta): -
return np.cos(3*np.pi*theta)/theta -
-
def minimaFunctionDerivative(theta): -
const1 = 3*np.pi -
const2 = const1*theta -
return -(const1*np.sin(const2)/theta)-np.cos(const2)/theta**2 -
-
theta = .6 -
iterations=45 -
learningRate = .0007 -
optimizedParameters = optimize(iterations,\ -
minimaFunction,\ -
minimaFunctionDerivative,\ -
theta,\ -
learningRate) -
这似乎运作得很好!您应该注意到,如果θ的初始值较大,则优化算法将在某一个局部极小处结束。然而,如上所述,在极高维度空间中这种可能性并不大,因为它要求所有参数同时满足凹函数。
你可能会想,「如果我们的学习率太大,会发生什么?」。如果步长过大,则算法可能永远不会找到如下的动画所示的最佳值。监控代价函数并确保它单调递减,这一点很重要。如果没有单调递减,可能需要降低学习率。
SGD 也适用于多变量参数空间的情况。我们可以将二维函数绘制成等高线图。在这里你可以看到 SGD 对一个不对称的碗形函数同样有效。
-
import numpy as np -
import matplotlib.mlab as mlab -
import matplotlib.pyplot as plt -
import scipy.stats -
import matplotlib.animation as animation -
-
def minimaFunction(params): -
#Bivariate Normal function -
X,Y = params -
-
sigma11,sigma12,mu11,mu12 = (3.0,.5,0.0,0.0) -
-
Z1 = mlab.bivariate_normal(X, Y, sigma11,sigma12,mu11,mu12) -
-
Z = Z1 -
-
return -40*Z -
-
def minimaFunctionDerivative(params): -
# Derivative of the bivariate normal function -
X,Y = params -
-
sigma11,sigma12,mu11,mu12 = (3.0,.5,0.0,0.0) -
-
dZ1X = -scipy.stats.norm.pdf(X, mu11, sigma11)*(mu11 - X)/sigma11**2 -
dZ1Y = -scipy.stats.norm.pdf(Y, mu12, sigma12)*(mu12 - Y)/sigma12**2 -
-
return (dZ1X,dZ1Y) -
-
def optimize(iterations, oF, dOF,params,learningRate,beta): -
""" -
computes the optimal value of params for a given objective function and its derivative -
Arguments: -
- iteratoins - the number of iterations required to optimize the objective function -
- oF - the objective function -
- dOF - the derivative function of the objective function -
- params - the parameters of the function to optimize -
- learningRate - the learning rate -
- beta - The weighted moving average parameter -
Return: -
- oParams - the list of optimized parameters at each step of iteration -
""" -
oParams = [params] -
vdw = (0.0,0.0) -
#The iteration loop -
for i in range(iterations): -
# Compute the derivative of the parameters -
dParams = dOF(params) -
-
#SGD in this line Goes through each parameter and applies parameter = parameter -learningrate*dParameter -
params = tuple([par-learningRate*dPar for dPar,par in zip(dParams,params)]) -
-
# append the new parameters -
oParams.append(params) -
-
return oParams -
-
iterations=100 -
learningRate = 1 -
beta = .9 -
x,y = 4.0,1.0 -
params = (x,y) -
optimizedParameters = optimize(iterations,\ -
minimaFunction,\ -
minimaFunctionDerivative,\ -
params,\ -
learningRate,\ -
beta) -
动量 SGD
注意,传统 SGD 没有解决所有问题!通常,用户想要使用非常大的学习速率来快速学习感兴趣的参数。不幸的是,当代价函数波动较大时,这可能导致不稳定。你可以看到,在前面的视频中,由于缺乏水平方向上的最小值,y 参数方向的抖动形式。动量算法试图使用过去的梯度预测学习率来解决这个问题。通常,使用动量的 SGD 通过以下公式更新参数:
γ 和 ν 值允许用户对 dJ(θ) 的前一个值和当前值进行加权来确定新的θ值。人们通常选择γ和ν的值来创建指数加权移动平均值,如下所示:
β参数的最佳选择是 0.9。选择一个等于 1-1/t 的β值可以让用户更愿意考虑νdw 的最新 t 值。这种简单的改变可以使优化过程产生显著的结果!我们现在可以使用更大的学习率,并在尽可能短的时间内收敛!
-
import numpy as np -
import matplotlib.mlab as mlab -
import matplotlib.pyplot as plt -
import scipy.stats -
import matplotlib.animation as animation -
-
def minimaFunction(params): -
#Bivariate Normal function -
X,Y = params -
-
sigma11,sigma12,mu11,mu12 = (3.0,.5,0.0,0.0) -
-
Z1 = mlab.bivariate_normal(X, Y, sigma11,sigma12,mu11,mu12) -
-
Z = Z1 -
-
return -40*Z -
-
def minimaFunctionDerivative(params): -
# Derivative of the bivariate normal function -
X,Y = params -
-
sigma11,sigma12,mu11,mu12 = (3.0,.5,0.0,0.0) -
-
dZ1X = -scipy.stats.norm.pdf(X, mu11, sigma11)*(mu11 - X)/sigma11**2 -
dZ1Y = -scipy.stats.norm.pdf(Y, mu12, sigma12)*(mu12 - Y)/sigma12**2 -
-
return (dZ1X,dZ1Y) -
-
def optimize(iterations, oF, dOF,params,learningRate,beta): -
""" -
computes the optimal value of params for a given objective function and its derivative -
Arguments: -
- iteratoins - the number of iterations required to optimize the objective function -
- oF - the objective function -
- dOF - the derivative function of the objective function -
- params - the parameters of the function to optimize -
- learningRate - the learning rate -
- beta - The weighted moving average parameter for momentum -
Return: -
- oParams - the list of optimized parameters at each step of iteration -
""" -
oParams = [params] -
vdw = (0.0,0.0) -
#The iteration loop -
for i in range(iterations): -
# Compute the derivative of the parameters -
dParams = dOF(params) -
-
# Compute the momentum of each gradient vdw = vdw*beta+(1.0+beta)*dPar -
vdw = tuple([vDW*beta+(1.0-beta)*dPar for dPar,vDW in zip(dParams,vdw)]) -
-
#SGD in this line Goes through each parameter and applies parameter = parameter -learningrate*dParameter -
params = tuple([par-learningRate*dPar for dPar,par in zip(vdw,params)]) -
-
# append the new parameters -
oParams.append(params) -
-
return oParams -
-
iterations=100 -
learningRate = 5.3 -
beta = .9 -
x,y = 4.0,1.0 -
params = (x,y) -
optimizedParameters = optimize(iterations,\ -
minimaFunction,\ -
minimaFunctionDerivative,\ -
params,\ -
learningRate,\ -
beta) -
RMSProp
像工程中的其它事物一样,我们一直在努力做得更好。RMS prop 试图通过观察关于每个参数的函数梯度的相对大小,来改善动量函数。因此,我们可以取每个梯度平方的加权指数移动平均值,并按比例归一化梯度下降函数。具有较大梯度的参数的 sdw 值将变得比具有较小梯度的参数大得多,从而使代价函数平滑下降到最小值。可以在下面的等式中看到:
请注意,这里的 epsilon 是为数值稳定性而添加的,可以取 10e-7。这是为什么昵?
-
import numpy as np -
import matplotlib.mlab as mlab -
import matplotlib.pyplot as plt -
import scipy.stats -
import matplotlib.animation as animation -
-
def minimaFunction(params): -
#Bivariate Normal function -
X,Y = params -
-
sigma11,sigma12,mu11,mu12 = (3.0,.5,0.0,0.0) -
-
Z1 = mlab.bivariate_normal(X, Y, sigma11,sigma12,mu11,mu12) -
-
Z = Z1 -
-
return -40*Z -
-
def minimaFunctionDerivative(params): -
# Derivative of the bivariate normal function -
X,Y = params -
-
sigma11,sigma12,mu11,mu12 = (3.0,.5,0.0,0.0) -
-
dZ1X = -scipy.stats.norm.pdf(X, mu11, sigma11)*(mu11 - X)/sigma11**2 -
dZ1Y = -scipy.stats.norm.pdf(Y, mu12, sigma12)*(mu12 - Y)/sigma12**2 -
-
return (dZ1X,dZ1Y) -
-
def optimize(iterations, oF, dOF,params,learningRate,beta): -
""" -
computes the optimal value of params for a given objective function and its derivative -
Arguments: -
- iteratoins - the number of iterations required to optimize the objective function -
- oF - the objective function -
- dOF - the derivative function of the objective function -
- params - the parameters of the function to optimize -
- learningRate - the learning rate -
- beta - The weighted moving average parameter for RMSProp -
Return: -
- oParams - the list of optimized parameters at each step of iteration -
""" -
oParams = [params] -
sdw = (0.0,0.0) -
eps = 10**(-7) -
#The iteration loop -
for i in range(iterations): -
# Compute the derivative of the parameters -
dParams = dOF(params) -
-
# Compute the momentum of each gradient sdw = sdw*beta+(1.0+beta)*dPar^2 -
sdw = tuple([sDW*beta+(1.0-beta)*dPar**2 for dPar,sDW in zip(dParams,sdw)]) -
-
#SGD in this line Goes through each parameter and applies parameter = parameter -learningrate*dParameter -
params = tuple([par-learningRate*dPar/((sDW**.5)+eps) for sDW,par,dPar in zip(sdw,params,dParams)]) -
-
# append the new parameters -
oParams.append(params) -
-
return oParams -
-
iterations=10 -
learningRate = .3 -
beta = .9 -
x,y = 5.0,1.0 -
params = (x,y) -
optimizedParameters = optimize(iterations,\ -
minimaFunction,\ -
minimaFunctionDerivative,\ -
params,\ -
learningRate,\ -
beta) -
Adam 算法
Adam 算法将动量和 RMSProp 的概念结合成一种算法,以获得两全其美的效果。公式如下:
-
import numpy as np -
import matplotlib.mlab as mlab -
import matplotlib.pyplot as plt -
import scipy.stats -
import matplotlib.animation as animation -
-
def minimaFunction(params): -
#Bivariate Normal function -
X,Y = params -
-
sigma11,sigma12,mu11,mu12 = (3.0,.5,0.0,0.0) -
-
Z1 = mlab.bivariate_normal(X, Y, sigma11,sigma12,mu11,mu12) -
-
Z = Z1 -
-
return -40*Z -
-
def minimaFunctionDerivative(params): -
# Derivative of the bivariate normal function -
X,Y = params -
-
sigma11,sigma12,mu11,mu12 = (3.0,.5,0.0,0.0) -
-
dZ1X = -scipy.stats.norm.pdf(X, mu11, sigma11)*(mu11 - X)/sigma11**2 -
dZ1Y = -scipy.stats.norm.pdf(Y, mu12, sigma12)*(mu12 - Y)/sigma12**2 -
-
return (dZ1X,dZ1Y) -
-
def optimize(iterations, oF, dOF,params,learningRate,beta1,beta2): -
""" -
computes the optimal value of params for a given objective function and its derivative -
Arguments: -
- iteratoins - the number of iterations required to optimize the objective function -
- oF - the objective function -
- dOF - the derivative function of the objective function -
- params - the parameters of the function to optimize -
- learningRate - the learning rate -
- beta1 - The weighted moving average parameter for momentum component of ADAM -
- beta2 - The weighted moving average parameter for RMSProp component of ADAM -
Return: -
- oParams - the list of optimized parameters at each step of iteration -
""" -
oParams = [params] -
vdw = (0.0,0.0) -
sdw = (0.0,0.0) -
vdwCorr = (0.0,0.0) -
sdwCorr = (0.0,0.0) -
-
eps = 10**(-7) -
#The iteration loop -
for i in range(iterations): -
# Compute the derivative of the parameters -
dParams = dOF(params) -
-
# Compute the momentum of each gradient vdw = vdw*beta+(1.0+beta)*dPar -
vdw = tuple([vDW*beta1+(1.0-beta1)*dPar for dPar,vDW in zip(dParams,vdw)]) -
-
# Compute the rms of each gradient sdw = sdw*beta+(1.0+beta)*dPar^2 -
sdw = tuple([sDW*beta2+(1.0-beta2)*dPar**2.0 for dPar,sDW in zip(dParams,sdw)]) -
# Compute the weight boosting for sdw and vdw -
vdwCorr = tuple([vDW/(1.0-beta1**(i+1.0)) for vDW in vdw]) -
sdwCorr = tuple([sDW/(1.0-beta2**(i+1.0)) for sDW in sdw]) -
-
#SGD in this line Goes through each parameter and applies parameter = parameter -learningrate*dParameter -
params = tuple([par-learningRate*vdwCORR/((sdwCORR**.5)+eps) for sdwCORR,vdwCORR,par in zip(vdwCorr,sdwCorr,params)]) -
-
# append the new parameters -
oParams.append(params) -
-
return oParams -
-
iterations=100 -
learningRate = .1 -
beta1 = .9 -
beta2 = .999 -
x,y = 5.0,1.0 -
params = (x,y) -
optimizedParameters = optimize(iterations,\ -
minimaFunction,\ -
minimaFunctionDerivative,\ -
params,\ -
learningRate,\ -
beta1,\ -
beta2)</div> -
Adam 算法可能是目前深度学习中使用最广泛的优化算法,适用于多种应用。Adam 计算了一个 νdw^corr 的值,用于加快指数加权移动平均值的变化。它将通过增加它们的值来对它们进行标准化,与当前的迭代次数成反比。使用 Adam 时有一些很好的初始值可供尝试。它最好以 0.9 的 β_1 和 0.999 的 β_2 开头。
总结
优化目标函数的算法有相当多的选择。在上述示例中,我们发现各种方法的收敛速度越来越快:
– SGD: 100 次迭代
– SGD+Momentum: 50 次迭代
– RMSProp: 10 次迭代
– ADAM: 5 次迭代
原文链接:https://3dbabove.com/2017/11/14/optimizationalgorithms/