温馨提示：

    本文将介绍统计学中的优化知识，凸优化和梯度下降，多为公式推导和图形化展示，较为硬核

优化与深度学习

优化与估计

尽管优化方法可以最小化深度学习中的损失函数值，但本质上优化方法达到的目标与深度学习的目标并不相同。

  • 优化方法目标：训练集损失函数值

  • 深度学习目标：测试集损失函数值（泛化性）

  • 借助图形直观比较

    1
    2
    3
    4
    5
    6
    %matplotlib inline
    import sys
    sys.path.append('path to file storge d2lzh1981')
    import d2lzh1981 as d2l
    from mpl_toolkits import mplot3d # 三维画图
    import numpy as np
1
2
3
4
5
6
7
8
9
10
11
12
def f(x): return x * np.cos(np.pi * x)
def g(x): return f(x) + 0.2 * np.cos(5 * np.pi * x)

d2l.set_figsize((5, 3))
x = np.arange(0.5, 1.5, 0.01)
fig_f, = d2l.plt.plot(x, f(x),label="train error")
fig_g, = d2l.plt.plot(x, g(x),'--', c='purple', label="test error")
fig_f.axes.annotate('empirical risk', (1.0, -1.2), (0.5, -1.1),arrowprops=dict(arrowstyle='->'))
fig_g.axes.annotate('expected risk', (1.1, -1.05), (0.95, -0.5),arrowprops=dict(arrowstyle='->'))
d2l.plt.xlabel('x')
d2l.plt.ylabel('risk')
d2l.plt.legend(loc="upper right")

优化在深度学习中的挑战

  1. 局部最小值
  2. 鞍点
  3. 梯度消失
局部最小值

$$
f(x) = x\cos \pi x
$$

1
2
3
4
5
6
7
8
9
10
11
12
def f(x):
    return x * np.cos(np.pi * x)

d2l.set_figsize((4.5, 2.5))
x = np.arange(-1.0, 2.0, 0.1)
fig,  = d2l.plt.plot(x, f(x))
fig.axes.annotate('local minimum', xy=(-0.3, -0.25), xytext=(-0.77, -1.0),
                  arrowprops=dict(arrowstyle='->'))
fig.axes.annotate('global minimum', xy=(1.1, -0.95), xytext=(0.6, 0.8),
                  arrowprops=dict(arrowstyle='->'))
d2l.plt.xlabel('x')
d2l.plt.ylabel('f(x)');
鞍点

函数在一阶导数为零处（驻点）的黑塞矩阵为不定矩阵。

1
2
3
4
5
6
x = np.arange(-2.0, 2.0, 0.1)
fig, = d2l.plt.plot(x, x**3)
fig.axes.annotate('saddle point', xy=(0, -0.2), xytext=(-0.52, -5.0),
                  arrowprops=dict(arrowstyle='->'))
d2l.plt.xlabel('x')
d2l.plt.ylabel('f(x)');

海森矩阵

$$
A=\left[\begin{array}{cccc}{\frac{\partial^{2} f}{\partial x_{1}^{2}}} & {\frac{\partial^{2} f}{\partial x_{1} \partial x_{2}}} & {\cdots} & {\frac{\partial^{2} f}{\partial x_{1} \partial x_{n}}} \\ {\frac{\partial^{2} f}{\partial x_{2} \partial x_{1}}} & {\frac{\partial^{2} f}{\partial x_{2}^{2}}} & {\cdots} & {\frac{\partial^{2} f}{\partial x_{2} \partial x_{n}}} \\ {\vdots} & {\vdots} & {\ddots} & {\vdots} \\ {\frac{\partial^{2} f}{\partial x_{n} \partial x_{1}}} & {\frac{\partial^{2} f}{\partial x_{n} \partial x_{2}}} & {\cdots} & {\frac{\partial^{2} f}{\partial x_{n}^{2}}}\end{array}\right]
$$

海森矩阵特征值和鞍点还有局部极小值的点的关系

偏导数为零的点

  • 特征值都大于零是局部极小值点
  • 都为负数是局部极大指点
  • 有正有负就是鞍点
1
2
3
4
5
6
7
8
9
10
11
12
13
x, y = np.mgrid[-1: 1: 31j, -1: 1: 31j]
z = x**2 - y**2

d2l.set_figsize((6, 4))
ax = d2l.plt.figure().add_subplot(111, projection='3d')
ax.plot_wireframe(x, y, z, **{'rstride': 2, 'cstride': 2})
ax.plot([0], [0], [0], 'ro', markersize=10)
ticks = [-1,  0, 1]
d2l.plt.xticks(ticks)
d2l.plt.yticks(ticks)
ax.set_zticks(ticks)
d2l.plt.xlabel('x')
d2l.plt.ylabel('y');
梯度消失
1
2
3
4
5
x = np.arange(-2.0, 5.0, 0.01)
fig, = d2l.plt.plot(x, np.tanh(x))
d2l.plt.xlabel('x')
d2l.plt.ylabel('f(x)')
fig.axes.annotate('vanishing gradient', (4, 1), (2, 0.0) ,arrowprops=dict(arrowstyle='->'))

梯度下降

1
2
3
4
5
6
7
8
9
%matplotlib inline
import numpy as np
import torch
import time
from torch import nn, optim
import math
import sys
sys.path.append('path to file storge d2lzh1981')
import d2lzh1981 as d2l

一维梯度下降

证明：沿梯度反方向移动自变量可以减小函数值

泰勒展开：

$$
f(x+\epsilon)=f(x)+\epsilon f^{\prime}(x)+\mathcal{O}\left(\epsilon^{2}\right)
$$

代入沿梯度方向的移动量 $\eta f^{\prime}(x)$：

$$
f\left(x-\eta f^{\prime}(x)\right)=f(x)-\eta f^{\prime 2}(x)+\mathcal{O}\left(\eta^{2} f^{\prime 2}(x)\right)
$$

$$
f\left(x-\eta f^{\prime}(x)\right) \lesssim f(x)
$$

$$
x \leftarrow x-\eta f^{\prime}(x)
$$

e.g.

$$
f(x) = x^2
$$

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
def f(x):
    return x**2  # Objective function

def gradf(x):
    return 2 * x  # Its derivative

def gd(eta):
    x = 10
    results = [x]
    for i in range(20):
        # eta 学习率
        x -= eta * gradf(x)
        results.append(x)
    print('epoch 20, x:', x)
    return results

res = gd(0.2)

梯度下降轨迹

1
2
3
4
5
6
7
8
9
10
11
def show_trace(res):
    n = max(abs(min(res)), abs(max(res)))
    f_line = np.arange(-n, n, 0.01)
    d2l.set_figsize((3.5, 2.5))
    d2l.plt.plot(f_line, [f(x) for x in f_line],'-')
    d2l.plt.plot(res, [f(x) for x in res],'-o')
    d2l.plt.xlabel('x')
    d2l.plt.ylabel('f(x)')
    

show_trace(res)
学习率

学习率过小 Code：show_trace(gd(0.05))
在这里插入图片描述

学习率过大 Code：show_trace(gd(1.1))
在这里插入图片描述

局部极小值

$$
f(x) = x\cos cx
$$

1
2
3
4
5
6
7
8
9
10
11
c = 0.15 * np.pi

def f(x):
    return x * np.cos(c * x)

def gradf(x):
    return np.cos(c * x) - c * x * np.sin(c * x)

# 学习率不合适容易导致
show_trace(gd(2))
show_trace(gd(0.5))

多维梯度下降

$$
\nabla f(\mathbf{x})=\left[\frac{\partial f(\mathbf{x})}{\partial x_{1}}, \frac{\partial f(\mathbf{x})}{\partial x_{2}}, \dots, \frac{\partial f(\mathbf{x})}{\partial x_{d}}\right]^{\top}
$$

$$
f(\mathbf{x}+\epsilon)=f(\mathbf{x})+\epsilon^{\top} \nabla f(\mathbf{x})+\mathcal{O}\left(|\epsilon|^{2}\right)
$$

$$
\mathbf{x} \leftarrow \mathbf{x}-\eta \nabla f(\mathbf{x})
$$

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
# 训练 trainer展示x如何更新
def train_2d(trainer, steps=20):
    x1, x2 = -5, -2
    results = [(x1, x2)]
    for i in range(steps):
        x1, x2 = trainer(x1, x2)
        results.append((x1, x2))
    print('epoch %d, x1 %f, x2 %f' % (i + 1, x1, x2))
    return results
# 垂直于等高线梯度下降
def show_trace_2d(f, results): 
    d2l.plt.plot(*zip(*results), '-o', color='#ff7f0e')
    x1, x2 = np.meshgrid(np.arange(-5.5, 1.0, 0.1), np.arange(-3.0, 1.0, 0.1))
    d2l.plt.contour(x1, x2, f(x1, x2), colors='#1f77b4')
    d2l.plt.xlabel('x1')
    d2l.plt.ylabel('x2')

e.g.

$$
f(x) = x_1^2 + 2x_2^2
$$

1
2
3
4
5
6
7
8
9
eta = 0.1

def f_2d(x1, x2):  # 目标函数
    return x1 ** 2 + 2 * x2 ** 2

def gd_2d(x1, x2):
    return (x1 - eta * 2 * x1, x2 - eta * 4 * x2)

show_trace_2d(f_2d, train_2d(gd_2d))

自适应方法

牛顿法

优势 :

梯度下降“步幅”的确定比较困难

而牛顿法相当于可以通过Hessian矩阵来调整“步幅”。

在牛顿法中，局部极小值也可以通过调整学习率来解决。

在 $x + \epsilon$ 处泰勒展开：

$$
f(\mathbf{x}+\epsilon)=f(\mathbf{x})+\epsilon^{\top} \nabla f(\mathbf{x})+\frac{1}{2} \epsilon^{\top} \nabla \nabla^{\top} f(\mathbf{x}) \epsilon+\mathcal{O}\left(|\epsilon|^{3}\right)
$$

最小值点处满足: $\nabla f(\mathbf{x})=0$, 即我们希望 $\nabla f(\mathbf{x} + \epsilon)=0$, 对上式关于 $\epsilon$ 求导，忽略高阶无穷小，有：

$$
\nabla f(\mathbf{x})+\boldsymbol{H}{f} \boldsymbol{\epsilon}=0 \text { and hence } \epsilon=-\boldsymbol{H}{f}^{-1} \nabla f(\mathbf{x})
$$

牛顿法需要计算Hessian矩阵的逆，计算量比较大。

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
c = 0.5

def f(x):
    return np.cosh(c * x)  # Objective

def gradf(x):
    return c * np.sinh(c * x)  # Derivative

def hessf(x):
    return c**2 * np.cosh(c * x)  # Hessian

# Hide learning rate for now
def newton(eta=1):
    x = 10
    results = [x]
    for i in range(10):
        x -= eta * gradf(x) / hessf(x)
        results.append(x)
    print('epoch 10, x:', x)
    return results

show_trace(newton())
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
# 牛顿法对于有局部极小值的情况
# 和梯度下降的方法有一样的效果
# 正确的方法还是降低学习率
c = 0.15 * np.pi

def f(x):
    return x * np.cos(c * x)

def gradf(x):
    return np.cos(c * x) - c * x * np.sin(c * x)

def hessf(x):
    return - 2 * c * np.sin(c * x) - x * c**2 * np.cos(c * x)

show_trace(newton())

show_trace(newton(0.5))
在这里插入图片描述

收敛性分析

只考虑在函数为凸函数, 且最小值点上 $f’’(x^*) >0$ 时的收敛速度：

令 $x_k$ 为第 $k$ 次迭代后 $x$ 的值， $e_{k}:=x_{k}-x^{*}$ 表示 $x_k$ 到最小值点 $x^{*}$ 的距离，由 $f’(x^{*}) = 0$:

$$
0=f^{\prime}\left(x_{k}-e_{k}\right)=f^{\prime}\left(x_{k}\right)-e_{k} f^{\prime \prime}\left(x_{k}\right)+\frac{1}{2} e_{k}^{2} f^{\prime \prime \prime}\left(\xi_{k}\right) \text{for some } \xi_{k} \in\left[x_{k}-e_{k}, x_{k}\right]
$$

两边除以 $f’’(x_k)$, 有：

$$
e_{k}-f^{\prime}\left(x_{k}\right) / f^{\prime \prime}\left(x_{k}\right)=\frac{1}{2} e_{k}^{2} f^{\prime \prime \prime}\left(\xi_{k}\right) / f^{\prime \prime}\left(x_{k}\right)
$$

代入更新方程 $x_{k+1} = x_{k} - f^{\prime}\left(x_{k}\right) / f^{\prime \prime}\left(x_{k}\right)$, 得到：

$$
x_k - x^{*} - f^{\prime}\left(x_{k}\right) / f^{\prime \prime}\left(x_{k}\right) =\frac{1}{2} e_{k}^{2} f^{\prime \prime \prime}\left(\xi_{k}\right) / f^{\prime \prime}\left(x_{k}\right)
$$

$$
x_{k+1} - x^{*} = e_{k+1} = \frac{1}{2} e_{k}^{2} f^{\prime \prime \prime}\left(\xi_{k}\right) / f^{\prime \prime}\left(x_{k}\right)
$$

当 $\frac{1}{2} f^{\prime \prime \prime}\left(\xi_{k}\right) / f^{\prime \prime}\left(x_{k}\right) \leq c$ 时，有:

$$
e_{k+1} \leq c e_{k}^{2}
$$

预处理 （Heissan阵辅助梯度下降）

$$
\mathbf{x} \leftarrow \mathbf{x}-\eta \operatorname{diag}\left(H_{f}\right)^{-1} \nabla \mathbf{x}
$$

梯度下降与线性搜索（共轭梯度法）

随机梯度下降

随机梯度下降参数更新

对于有 $n$ 个样本对训练数据集，设 $f_i(x)$ 是第 $i$ 个样本的损失函数, 则目标函数为:

$$
f(\mathbf{x})=\frac{1}{n} \sum_{i=1}^{n} f_{i}(\mathbf{x})
$$

其梯度为:

$$
\nabla f(\mathbf{x})=\frac{1}{n} \sum_{i=1}^{n} \nabla f_{i}(\mathbf{x})
$$

每一个样本的梯度是对整体的梯度的无偏估计

使用该梯度的一次更新的时间复杂度为 $\mathcal{O}(n)$

随机梯度下降更新公式 $\mathcal{O}(1)$:

$$
\mathbf{x} \leftarrow \mathbf{x}-\eta \nabla f_{i}(\mathbf{x})
$$

且有：

$$
\mathbb{E}{i} \nabla f{i}(\mathbf{x})=\frac{1}{n} \sum_{i=1}^{n} \nabla f_{i}(\mathbf{x})=\nabla f(\mathbf{x})
$$
e.g.

$$
f(x_1, x_2) = x_1^2 + 2 x_2^2
$$

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
def f(x1, x2):
    return x1 ** 2 + 2 * x2 ** 2  # Objective

def gradf(x1, x2):
    return (2 * x1, 4 * x2)  # Gradient

def sgd(x1, x2):  # Simulate noisy gradient
    global lr  # Learning rate scheduler
    (g1, g2) = gradf(x1, x2)  # Compute gradient
    (g1, g2) = (g1 + np.random.normal(0.1), g2 + np.random.normal(0.1))
    eta_t = eta * lr()  # Learning rate at time t
    return (x1 - eta_t * g1, x2 - eta_t * g2)  # Update variables

eta = 0.1
lr = (lambda: 1)  # Constant learning rate
show_trace_2d(f, train_2d(sgd, steps=50))

动态学习率

  • 在最开始学习率设计比较大，加速收敛
  • 学习率可以设计为指数衰减或多项式衰减
  • 在优化进行一段时间后可以适当减小学习率来避免振荡

$$
\begin{array}{ll}{\eta(t)=\eta_{i} \text { if } t_{i} \leq t \leq t_{i+1}} & {\text { piecewise constant }} \\ {\eta(t)=\eta_{0} \cdot e^{-\lambda t}} & {\text { exponential }} \\ {\eta(t)=\eta_{0} \cdot(\beta t+1)^{-\alpha}} & {\text { polynomial }}\end{array}
$$

1
2
3
4
5
6
7
8
def exponential():
    global ctr
    ctr += 1
    return math.exp(-0.1 * ctr)

ctr = 1
lr = exponential  # Set up learning rate
show_trace_2d(f, train_2d(sgd, steps=1000))
1
2
3
4
5
6
7
8
9
def polynomial():
    # 迭代次数
    global ctr
    ctr += 1
    return (1 + 0.1 * ctr)**(-0.5)

ctr = 1
lr = polynomial  # Set up learning rate
show_trace_2d(f, train_2d(sgd, steps=50))

小批量随机梯度下降

读取数据

读取数据

1
2
3
4
5
6
7
8
def get_data_ch7():
    data = np.genfromtxt('/home/kesci/input/airfoil4755/airfoil_self_noise.dat', delimiter='\t')
    data = (data - data.mean(axis=0)) / data.std(axis=0) # 标准化
    return torch.tensor(data[:1500, :-1], dtype=torch.float32), \
           torch.tensor(data[:1500, -1], dtype=torch.float32) # 前1500个样本(每个样本5个特征)

features, labels = get_data_ch7()
features.shape

数据可视化

1
2
3
import pandas as pd
df = pd.read_csv('path to airfoil_self_noise.dat', delimiter='\t', header=None)
df.head(10)

Stochastic Gradient Descent (SGD)函数

1
2
3
def sgd(params, states, hyperparams):
    for p in params:
        p.data -= hyperparams['lr'] * p.grad.data

训练

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
def train_ch7(optimizer_fn, states, hyperparams, features, labels,
              batch_size=10, num_epochs=2):
    # 初始化模型
    net, loss = d2l.linreg, d2l.squared_loss
    
    w = torch.nn.Parameter(torch.tensor(np.random.normal(0, 0.01, size=(features.shape[1], 1)), dtype=torch.float32),
                           requires_grad=True)
    b = torch.nn.Parameter(torch.zeros(1, dtype=torch.float32), requires_grad=True)

    def eval_loss():
        return loss(net(features, w, b), labels).mean().item()

    ls = [eval_loss()]
    data_iter = torch.utils.data.DataLoader(
        torch.utils.data.TensorDataset(features, labels), batch_size, shuffle=True)
    
    for _ in range(num_epochs):
        start = time.time()
        for batch_i, (X, y) in enumerate(data_iter):
            l = loss(net(X, w, b), y).mean()  # 使用平均损失
            
            # 梯度清零
            if w.grad is not None:
                w.grad.data.zero_()
                b.grad.data.zero_()
                
            l.backward()
            optimizer_fn([w, b], states, hyperparams)  # 迭代模型参数
            if (batch_i + 1) * batch_size % 100 == 0:
                ls.append(eval_loss())  # 每100个样本记录下当前训练误差
    # 打印结果和作图
    print('loss: %f, %f sec per epoch' % (ls[-1], time.time() - start))
    d2l.set_figsize()
    d2l.plt.plot(np.linspace(0, num_epochs, len(ls)), ls)
    d2l.plt.xlabel('epoch')
    d2l.plt.ylabel('loss')

测试

1
2
def train_sgd(lr, batch_size, num_epochs=2):
    train_ch7(sgd, None, {'lr': lr}, features, labels, batch_size, num_epochs)

Result

  • train_sgd(1, 1500, 6)
    在这里插入图片描述
  • train_sgd(0.005, 1)
    在这里插入图片描述
  • train_sgd(0.05, 10)
    在这里插入图片描述

简化模型

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
def train_pytorch_ch7(optimizer_fn, optimizer_hyperparams, features, labels,
                    batch_size=10, num_epochs=2):
    # 初始化模型
    net = nn.Sequential(
        nn.Linear(features.shape[-1], 1)
    )
    loss = nn.MSELoss()
    optimizer = optimizer_fn(net.parameters(), **optimizer_hyperparams)

    def eval_loss():
        return loss(net(features).view(-1), labels).item() / 2

    ls = [eval_loss()]
    data_iter = torch.utils.data.DataLoader(
        torch.utils.data.TensorDataset(features, labels), batch_size, shuffle=True)

    for _ in range(num_epochs):
        start = time.time()
        for batch_i, (X, y) in enumerate(data_iter):
            # 除以2是为了和train_ch7保持一致, 因为squared_loss中除了2
            l = loss(net(X).view(-1), y) / 2 
            
            optimizer.zero_grad()
            l.backward()
            optimizer.step()
            if (batch_i + 1) * batch_size % 100 == 0:
                ls.append(eval_loss())
    # 打印结果和作图
    print('loss: %f, %f sec per epoch' % (ls[-1], time.time() - start))
    d2l.set_figsize()
    d2l.plt.plot(np.linspace(0, num_epochs, len(ls)), ls)
    d2l.plt.xlabel('epoch')
    d2l.plt.ylabel('loss')

train_pytorch_ch7(optim.SGD, {“lr”: 0.05}, features, labels, 10)
在这里插入图片描述