Cs231n作业：Q1-2 Support Vector Machine（SVM）
SVM算法
略…
步骤
- 1.定义一个损失函数，量化我们对训练数据中分数的不满意程度。
- 2.提出一种有效地找出使损失函数最小的参数的方法。
SVM损失：计算了所有不正确的例子，将所有不正确类别的评分与正确类别评分之差再加1，将得到的数值与0比较，取二者最大，然后将所有数值进行求和。
计算分数：
s
=
f
(
x
,
W
)
=
W
x
s=f(x,W) =Wx
s=f(x,W)=Wx
计算完全损失（有正则项）：
L
=
1
N
∑
i
=
1
N
∑
j
≠
y
i
m
a
x
(
0
,
f
(
x
i
;
W
)
)
j
−
f
(
x
i
;
W
)
y
i
+
1
)
+
λ
R
(
W
)
L=\frac{ 1}{N}\sum_{i=1}^{N}\sum_{j≠y_i}^{}max(0,f(x_i; W))_j-f(x_i; W)y_i + 1) + λR(W)
L=N1i=1∑Nj̸=yi∑max(0,f(xi;W))j−f(xi;W)yi+1)+λR(W)
R
(
W
)
=
∑
k
∑
l
W
k
,
l
2
R(W) = \sum_{k}\sum_{l}W_k,_l^2
R(W)=k∑l∑Wk,l2
梯度计算（数值方法）：
由
L
i
=
∑
j
≠
y
i
m
a
x
(
0
,
W
j
X
i
T
−
W
y
i
X
i
T
+
1
)
，
也
即
L
i
=
∑
j
≠
y
i
m
a
x
(
0
,
S
j
−
S
y
i
+
1
)
L_i=\sum_{j≠y_i}max(0, W_jX_i^T-Wy_iX_i^T+1)，也即L_i=\sum_{j≠y_i}max(0, S_j-Sy_i+1)
Li=j̸=yi∑max(0,WjXiT−WyiXiT+1)，也即Li=j̸=yi∑max(0,Sj−Syi+1)知，当Li <= 0时，梯度为0，只有当Li > 0其梯度为：
当
j
≠
y
i
:
∂
L
i
∂
S
j
=
X
i
T
当j≠y_i : \frac{\partial L_i}{\partial S_j}=X_i^T
当j̸=yi:∂Sj∂Li=XiT
当
j
＝
y
i
:
∂
L
i
∂
S
y
i
=
−
X
i
T
当j＝y_i : \frac{\partial L_i}{\partial Sy_i}=-X_i^T
当j＝yi:∂Syi∂Li=−XiT
最后计算平均值以及加入正则化。
所以再在下面svm_loss_naive代码中：
if margin > 0:
dW[:,y[i]] += -X[i,:]
dW[:,j] += X[i,:]
dW /= num_train # 获取平均
dW += reg * W # 加入正则化
例子
1.计算分数及对应的损失：
2.计算平均损失：
3.加入正则项
Cs231n——SVM作业
Multiclass Support Vector Machine exercise
在这个练习中，你会:
• 为SVM实现全矢量化的损失函数
• 实现其解析梯度的全矢量表达式
• 使用数值梯度检查实现
• 使用验证集来调整学习速度和正则化强度
• 使用SGD优化损失函数
• 想象最终学习到的重量
CIFAR-10 Data Loading and Preprocessing
这里的数据集用的是CIFAR-10，数据的加载与预处理
# Load the raw CIFAR-10 data.
cifar10_dir = 'cs231n/datasets/cifar-10-batches-py'
# 清理变量以防止多次加载数据(这可能会导致内存问题)
try:
del X_train, y_train
del X_test, y_test
print('Clear previously loaded data.')
except:
pass
X_train, y_train, X_test, y_test = load_CIFAR10(cifar10_dir)
# As a sanity check, we print out the size of the training and test data.
print('Training data shape: ', X_train.shape)
print('Training labels shape: ', y_train.shape)
print('Test data shape: ', X_test.shape)
print('Test labels shape: ', y_test.shape)
通过训练数据和测试数据的大小可知：每张图片像素都是32 x 32 x 3，训练集有50000张，测试集有10000张。
输出：
Training data shape: (50000, 32, 32, 3)
Training labels shape: (50000,)
Test data shape: (10000, 32, 32, 3)
Test labels shape: (10000,)
这里展示来自每个类的一些图片例子：
# 从数据集中可视化一些示例。
# 我们展示了来自每个类的一些训练图像的例子。
classes = ['plane', 'car', 'bird', 'cat', 'deer', 'dog', 'frog', 'horse', 'ship', 'truck']
num_classes = len(classes)
samples_per_class = 7
for y, cls in enumerate(classes):
idxs = np.flatnonzero(y_train == y)
idxs = np.random.choice(idxs, samples_per_class, replace=False)
for i, idx in enumerate(idxs):
plt_idx = i * num_classes + y + 1
plt.subplot(samples_per_class, num_classes, plt_idx)
plt.imshow(X_train[idx].astype('uint8'))
plt.axis('off')
if i == 0:
plt.title(cls)
plt.show()
输出：
为了更有效的执行代码，将数据分割为训练集，验证集和测试集。（把训练集的前49000作为训练集，后1000作为验证集，选测试集中的前1000张作为测试集）。此外，我们将创建一个小的开发集作为训练数据的子集，可以将其用于开发，从而使代码运行得更快：
num_training = 49000
num_validation = 1000
num_test = 1000
num_dev = 500
# Our validation set will be num_validation points from the original
# training set.
# 将数据集中最后1000个数据作为验证集
mask = range(num_training, num_training + num_validation)
X_val = X_train[mask]
y_val = y_train[mask]
# Our training set will be the first num_train points from the original
# training set.
# 将数据集前49000个数据作为训练集
mask = range(num_training)
X_train = X_train[mask]
y_train = y_train[mask]
# We will also make a development set, which is a small subset of
# the training set.
# 在数据集中从0-49000数字中随机抽取num_dev大小的数据作为开发集，并且不能重用元素
mask = np.random.choice(num_training, num_dev, replace=False)
X_dev = X_train[mask]
y_dev = y_train[mask]
# We use the first num_test points of the original test set as our
# test set.
# 选测试集中的前1000个作为测试集。
mask = range(num_test)
X_test = X_test[mask]
y_test = y_test[mask]
print('Train data shape: ', X_train.shape)
print('Train labels shape: ', y_train.shape)
print('Validation data shape: ', X_val.shape)
print('Validation labels shape: ', y_val.shape)
print('Test data shape: ', X_test.shape)
print('Test labels shape: ', y_test.shape)
输出：
训练集有49000张，验证集有1000张，测试集有1000张。
Train data shape: (49000, 32, 32, 3)
Train labels shape: (49000,)
Validation data shape: (1000, 32, 32, 3)
Validation labels shape: (1000,)
Test data shape: (1000, 32, 32, 3)
Test labels shape: (1000,)
预处理：将图片数据进行张量变形，重新塑成一行
# Preprocessing: reshape the image data into rows
X_train = np.reshape(X_train, (X_train.shape[0], -1))
X_val = np.reshape(X_val, (X_val.shape[0], -1))
X_test = np.reshape(X_test, (X_test.shape[0], -1))
X_dev = np.reshape(X_dev, (X_dev.shape[0], -1))
# As a sanity check, print out the shapes of the data
# 作为一个完整性检查，打印出数据的形状
print('Training data shape: ', X_train.shape)
print('Validation data shape: ', X_val.shape)
print('Test data shape: ', X_test.shape)
print('dev data shape: ', X_dev.shape)
输出：
以X_train.shape为例：第一维大小为X_train.shape[0]即变为49000 而第二维为-1表示列不知道多少，所以根据剩下纬度进行计算，即32x32x3=3027。所以最终形状为(49000,3272)。
所以，作为一个完整性检查，打印出数据的形状：
Training data shape: (49000, 3072)
Validation data shape: (1000, 3072)
Test data shape: (1000, 3072)
dev data shape: (500, 3072)
进一步预处理：减去图像的均值
首先，根据训练数据计算图像均值
# first: compute the image mean based on the training data
mean_image = np.mean(X_train, axis=0) # 压缩行，对各列进行求均值
print(mean_image[:10]) # print a few of the elements
plt.figure(figsize=(4,4))
plt.imshow(mean_image.reshape((32,32,3)).astype('uint8')) # visualize the mean image
plt.show()
输出：
[130.64189796 135.98173469 132.47391837 130.05569388 135.34804082
131.75402041 130.96055102 136.14328571 132.47636735 131.48467347]
第二，从训练和测试数据中减去平均图像
# second: subtract the mean image from train and test data
X_train -= mean_image
X_val -= mean_image
X_test -= mean_image
X_dev -= mean_image
第三，增加一维偏置(即偏置技巧)，使得SVM只需要优化一个权值矩阵W。
# third: append the bias dimension of ones (i.e. bias trick) so that our SVM
# only has to worry about optimizing a single weight matrix W.
# 水平堆叠序列中的数组（列方向），即最最后一列后面增加一列np.ones((X_train.shape[0], 1))（作为偏置）
X_train = np.hstack([X_train, np.ones((X_train.shape[0], 1))])
X_val = np.hstack([X_val, np.ones((X_val.shape[0], 1))])
X_test = np.hstack([X_test, np.ones((X_test.shape[0], 1))])
X_dev = np.hstack([X_dev, np.ones((X_dev.shape[0], 1))])
print(X_train.shape, X_val.shape, X_test.shape, X_dev.shape)
输出：
(49000, 3073) (1000, 3073) (1000, 3073) (500, 3073)
SVM Classifier
本节的代码将全部在cs231n/classifier /linear_svm.py中编写。
实现简单（带循环）的结构化SVM损失函数：
输入有D维即特征，C个类别，我们操作是的N个小批量列子。
输入：
- W：包含权重的形状为(D, C)的数组
- X：包含少量数据的形状为(N, D)的数组
- y：包含训练标签，形状为(N, )的数组，其中y[i] = c 表示X[i]有标签c，其中 0<= c < C
- reg:(float)正则化强度
返回一个元组：
- loss as single float
- 对权值W的梯度;与W形状相同的数组
from builtins import range
import numpy as np
from random import shuffle
from past.builtins import xrange
def svm_loss_naive(W, X, y, reg):
"""
Structured SVM loss function, naive implementation (with loops).
Inputs have dimension D, there are C classes, and we operate on minibatches
of N examples.
Inputs:
- W: A numpy array of shape (D, C) containing weights.
- X: A numpy array of shape (N, D) containing a minibatch of data.
- y: A numpy array of shape (N,) containing training labels; y[i] = c means
that X[i] has label c, where 0 <= c < C.
- reg: (float) regularization strength
Returns a tuple of:
- loss as single float
- gradient with respect to weights W; an array of same shape as W
"""
# 初始化梯度为0。 (D, C)
dW = np.zeros(W.shape) # initialize the gradient as zero
# compute the loss and the gradient 计算损失和梯度
num_classes = W.shape[1] # 取对应类别 C
num_train = X.shape[0] # 取对应样本数 N
loss = 0.0
for i in range(num_train): # 遍历样本i-N
scores = X[i].dot(W) # 分别计算分数向量(1xC)，scores vecotr: s = f(xi,W)
correct_class_score = scores[y[i]] # 对应次样本真正标签所对应的分数(1X1)
for j in range(num_classes): # 遍历类别j-C
if j == y[i]: # 如果当前类别即为本样本标签，则跳过
continue
# 否则计算该类别 the SVM loss,注意 delta = 1,j≠y_i的均通过S_j - S_yi + 1 分别进行计算。
margin = scores[j] - correct_class_score + 1 # 获取对应一个实数
if margin > 0: # max函数在括号里≤0时，梯度肯定为0（初始化的值），所以直接看＞0
loss += margin # 该样本的损失等于该样本所得到的间隔实数
# 计算梯度：对W求偏导
# (X_iW_j - X_iW_yi + 1)对W_yi这列，通过减去X_i
# 所以从dW中取该类真正标签类别的所有特征[:,y[i]]（此时全为0），使其减去该类别所有特征值
dW[:,y[i]] += -X[i,:]
# (X_iW_j - X_iW_yi + 1)对W_j这列，通过加X_i
# 所以从dW中分别取出非真正标签类别的所有特征[:,j]（此时也全为0），使其加上该类别所有特征值。
dW[:,j] += X[i,:]
# Right now the loss is a sum over all training examples, but we want it
# to be an average instead so we divide by num_train.
loss /= num_train # 获取该样本的平均损失
dW /= num_train # 获取平均
# Add regularization to the loss.
loss += reg * np.sum(W * W) # 加入正则化，得到完整的损失函数
dW += reg * W # 加入正则化
#############################################################################
# TODO: #
# Compute the gradient of the loss function and store it dW. #
# Rather that first computing the loss and then computing the derivative, #
# it may be simpler to compute the derivative at the same time that the #
# loss is being computed. As a result you may need to modify some of the #
# code above to compute the gradient. #
#############################################################################
# *****START OF YOUR CODE (DO NOT DELETE/MODIFY THIS LINE)*****
# *****END OF YOUR CODE (DO NOT DELETE/MODIFY THIS LINE)*****
return loss, dW
如您所见，我们已经预填充了compute_loss_naive函数，该函数使用for循环来评估多类SVM的损失函数。
# 评估我们为您提供的损失的简单实现
from cs231n.classifiers.linear_svm import svm_loss_naive
import time
# 生成一个随机数小的SVM权值矩阵
W = np.random.randn(3073, 10) * 0.0001
loss, grad = svm_loss_naive(W, X_dev, y_dev, 0.000005)
print('loss: %f' % (loss, ))
输出：
loss: 9.112255
推导并实现SVM代价函数的梯度，并在函数svm_loss_naive中内联实现梯度。您会发现在现有函数中插入新代码很有帮助。
要检查是否正确地实现了梯度，可以用数值方法估计损失函数的梯度，并将数值估计与计算的梯度进行比较。我们已经为您提供了这样做的代码:
# 实现梯度之后，使用下面的代码重新计算梯度
# 用我们提供的函数来检查梯度
# 计算损失及其在W处的梯度.
loss, grad = svm_loss_naive(W, X_dev, y_dev, 0.0)
# 沿随机选择的几个维度数值计算梯度
# 将它们与分析计算的梯度进行比较。
# 数字应该匹配几乎沿着所有的维度。
from cs231n.gradient_check import grad_check_sparse
f = lambda w: svm_loss_naive(w, X_dev, y_dev, 0.0)[0]
grad_numerical = grad_check_sparse(f, W, grad)
# 当正则化打开时，是否再次检查梯度
# you didn't forget the regularization gradient did you?
loss, grad = svm_loss_naive(W, X_dev, y_dev, 5e1)
f = lambda w: svm_loss_naive(w, X_dev, y_dev, 5e1)[0]
grad_numerical = grad_check_sparse(f, W, grad)
输入，即检查numeric gradient和analytic gradient是否相同：
numerical: 16.558587 analytic: 16.558587, relative error: 7.044564e-12
numerical: -1.877586 analytic: -1.877586, relative error: 7.754054e-11
numerical: -17.992739 analytic: -17.992739, relative error: 9.499249e-12
numerical: 26.182227 analytic: 26.182227, relative error: 9.835010e-12
numerical: -54.546606 analytic: -54.546606, relative error: 1.438603e-11
numerical: 18.977124 analytic: 18.977124, relative error: 2.583222e-12
numerical: 20.062216 analytic: 20.062216, relative error: 1.051530e-12
numerical: 18.542379 analytic: 18.542379, relative error: 5.079020e-12
numerical: 26.232349 analytic: 26.232349, relative error: 6.182360e-12
numerical: -31.468373 analytic: -31.468373, relative error: 1.159591e-11
numerical: -48.611978 analytic: -48.617267, relative error: 5.440557e-05
numerical: 1.501071 analytic: 1.508248, relative error: 2.385005e-03
numerical: -7.296033 analytic: -7.294829, relative error: 8.252591e-05
numerical: -0.337898 analytic: -0.347751, relative error: 1.437046e-02
numerical: 24.106606 analytic: 24.120138, relative error: 2.806025e-04
numerical: 3.226319 analytic: 3.227376, relative error: 1.636281e-04
numerical: 36.533221 analytic: 36.539594, relative error: 8.722431e-05
numerical: -19.570135 analytic: -19.564052, relative error: 1.554484e-04
numerical: -40.827530 analytic: -40.827961, relative error: 5.273252e-06
numerical: -10.187518 analytic: -10.186827, relative error: 3.389178e-05
Inline Question1
有时，gradcheck中的维度可能并不完全匹配。造成这种差异的原因是什么呢?这是担忧的原因吗?在一维中，梯度检查可能失败的简单例子是什么?如何改变这种情况发生频率的边际效应?提示:SVM的损失函数严格来说不是可微的
Your Answier： 因为SVM的损失函数严格来说不是可微的。
接下来实现svm_loss_vectorized函数;现在只计算损失;我们稍后将实现梯度。
def svm_loss_vectorized(W, X, y, reg):
"""
Structured SVM loss function, vectorized implementation.
Inputs and outputs are the same as svm_loss_naive.
"""
loss = 0.0
dW = np.zeros(W.shape) # initialize the gradient as zero
scores = X.dot(W) # N*C的矩阵
num_train = X.shape[0]
#############################################################################
# TODO: #
# Implement a vectorized version of the structured SVM loss, storing the #
# result in loss. #
#############################################################################
# *****START OF YOUR CODE (DO NOT DELETE/MODIFY THIS LINE)*****
#第一个参数表示取行的范围，np.arange(num_train)=500，即取所有行（总共行为500）
#第二个参数表示取列。
# 所以就是取0行的多少列，1行的多少列，2行的多少列， 最终得到每张图片，正确标签对应的分数。
correct_scores = scores[np.arange(num_train),y] # 1xN
correct_scores = correct_scores.reshape((num_train, -1)) # Nx1
margins = np.maximum(0,scores - correct_scores + 1) # 计算误差 NxC
margins[range(num_train), y] = 0 # 将label值所在的位置误差置零
loss+=np.sum(margins)
loss/=num_train # 取所有损失记录结果平均值
loss+=reg*np.sum(W*W) # 加上正则化
# *****END OF YOUR CODE (DO NOT DELETE/MODIFY THIS LINE)*****
#############################################################################
# TODO: #
# Implement a vectorized version of the gradient for the structured SVM #
# loss, storing the result in dW. #
# #
# Hint: Instead of computing the gradient from scratch, it may be easier #
# to reuse some of the intermediate values that you used to compute the #
# loss. #
#############################################################################
# *****START OF YOUR CODE (DO NOT DELETE/MODIFY THIS LINE)*****
# 将margins>0的项（有误差的项）置为1，没误差的项为0
margins[margins > 0] = 1 # NxC
# 没误差的项中有一项为标记项，计算标记项的权重分量对误差也有共享，也需要更新对应的权重分量
# margins中这个参数就是当前样本结果错误分类的数量
row_num = -np.sum(margins,1)
margins[np.arange(num_train), y] = row_num
# X: 200x3073 margins:200x10 -> 10x3072
dW += np.dot(X.T, margins) # 3073x10
dW /= num_train # 平均权重
dW += reg * W # 正则化
# *****END OF YOUR CODE (DO NOT DELETE/MODIFY THIS LINE)*****
return loss, dW
# 接下来实现svm_loss_vectorized函数;现在只计算损失;我们稍后将实现梯度。
tic = time.time()
loss_naive, grad_naive = svm_loss_naive(W, X_dev, y_dev, 0.000005)
toc = time.time()
print('Naive loss: %e computed in %fs' % (loss_naive, toc - tic))
from cs231n.classifiers.linear_svm import svm_loss_vectorized
tic = time.time()
loss_vectorized, _ = svm_loss_vectorized(W, X_dev, y_dev, 0.000005)
toc = time.time()
print('Vectorized loss: %e computed in %fs' % (loss_vectorized, toc - tic))
# The losses should match but your vectorized implementation should be much faster.
print('difference: %f' % (loss_naive - loss_vectorized))
输出：
Naive loss: 9.112255e+00 computed in 0.157111s
Vectorized loss: 9.112255e+00 computed in 0.004004s
difference: -0.000000
# 完成svm_loss_vectorized的实现，并计算梯度的损失函数，以矢量化的方式。
# 初始实现和向量化实现应该匹配，
# 但是矢量化的版本应该会更快。
tic = time.time()
_, grad_naive = svm_loss_naive(W, X_dev, y_dev, 0.000005)
toc = time.time()
print('Naive loss and gradient: computed in %fs' % (toc - tic))
tic = time.time()
_, grad_vectorized = svm_loss_vectorized(W, X_dev, y_dev, 0.000005)
toc = time.time()
print('Vectorized loss and gradient: computed in %fs' % (toc - tic))
# 损失是一个数字，因此比较这两种实现计算的值很容易。
# 另一方面梯度是一个矩阵，所以我们用弗洛贝尼乌斯范数来比较它们。
difference = np.linalg.norm(grad_naive - grad_vectorized, ord='fro')
print('difference: %f' % difference)
输入：
Naive loss and gradient: computed in 0.150105s
Vectorized loss and gradient: computed in 0.004003s
difference: 0.000000
Stochastic Gradient Descent（随机梯度下降）
在一个训练循环内，必要时一直重复这些步骤：
（1） 抽取训练样本x和对应目标y组成的数据批量。
（2） 在x上运行网络[这一步叫作前向传播]，得到预测值y_pred。
（3） 计算网络在这批数据上的损失，用于衡量y_pred和y之间的距离。
（4） 更新网络的所有权重，使网络在这批数据上的损失略微下降。
最终得到的网络在训练数据上的损失非常小，即预测值y_pred和预期目标y之间的距离非常小。
基于当前在随机数据批量上的损失，一点一点地对参数进行调节。由于处理的是一个可微函数，你可以计算出它的梯度，从而有效地实现第四步。沿着梯度的反方向更新权重，损失每次都会变小一点。
步骤
1.抽取训练样本x和对应目标y组成的数据批量。
2.在x上运行网络，得到预测值y_pred。
3.计算网络在这批数据上的损失，用于衡量y_pred和y之间的距离。
4.计算损失相对于网络参数的梯度[一次反向传播]。
5.将参数沿着梯度的反方向移动一点，比如W -= step * gradient，从而使这批数据上的损失减小一点。
我们现在有了矢量化的有效的损失表达式，梯度和我们的梯度匹配的数值梯度。因此，我们准备做SGD以减少损失：
在文件linear_classifier中，在函数中实现SGD：
def train(self, X, y, learning_rate=1e-3, reg=1e-5, num_iters=100,
batch_size=200, verbose=False):
"""
Train this linear classifier using stochastic gradient descent.
训练这个线性分类器使用随机梯度下降。
Inputs:
- X: 包含训练数据的形状(N, D)的numpy数组;有N个每个维度D的训练样本。
- y: 包含训练标签的形状(N，)的numpy数组;y[i]= c表示X[i]对于c类有标签 0 <= c < C for C classes.
- learning_rate: (float)用于优化的学习率。
- reg: (float)正则化强度。
- num_iters: (整数)优化时要采取的步骤数
- batch_size: (整数)在每个步骤中使用的训练示例的数量。
- verbose: (boolean)如果为真，则在优化期间打印进度。
Outputs:
包含每次训练迭代时损失函数值的列表。
"""
num_train, dim = X.shape # 分别获取样本数量，以及特征数（纬度）
num_classes = np.max(y) + 1 # 获取类的个数， 假设y取0…K-1，其中K是类的个数
if self.W is None:
# 延迟初始化W
self.W = 0.001 * np.random.randn(dim, num_classes)
# 运行随机梯度下降来优化W
loss_history = []
for it in range(num_iters): # 遍历，(整数)优化时要采取的步骤数
X_batch = None
y_batch = None
#########################################################################
# TODO: #
# 从训练数据及其对应的标签中提取batch_size元素样本，用于这一轮梯度下降 #
# 将数据存储在X_batch中，相应的标签存储在y_batch中; #
# 采样后X_batch应该有shape (batch_size, dim)， #
# y_batch应该有shape (batch_size，) #
# Hint: Use np.random.choice to generate indices. Sampling with #
# replacement is faster than sampling without replacement. #
#########################################################################
# *****START OF YOUR CODE (DO NOT DELETE/MODIFY THIS LINE)*****
i = np.random.choice(a=num_train, size=batch_size) # 取num_train中，随机选取大小为batch_size的数据
X_batch = X[i,:] # 获取所选取的i个样本，及其对应的特征
y_batch = y[i] # 获取所选取的i个样本的类标签
# *****END OF YOUR CODE (DO NOT DELETE/MODIFY THIS LINE)*****
# 评估损失和梯度
loss, grad = self.loss(X_batch, y_batch, reg)
loss_history.append(loss)
# 执行参数更新
#########################################################################
# TODO: #
# Update the weights using the gradient and the learning rate. #
# 使用梯度和学习率更新权重。 #
#########################################################################
# *****START OF YOUR CODE (DO NOT DELETE/MODIFY THIS LINE)*****
# 将参数沿着梯度的反方向移动一点，从而使这批数据上的损失减小一点
# learning_rate 是步长（学习率），grad是梯度
self.W -= learning_rate*grad
# *****END OF YOUR CODE (DO NOT DELETE/MODIFY THIS LINE)*****
if verbose and it % 100 == 0:
print('iteration %d / %d: loss %f' % (it, num_iters, loss))
return loss_history
def loss(self, X_batch, y_batch, reg):
"""
计算损失函数及其导数
Compute the loss function and its derivative..
子类将覆盖它
Subclasses will override this.
Inputs:
- X_batch:形状(N, D)的numpy数组，包含N个数据点;每个点都有维数D。
data points; each point has dimension D.
- y_batch: 一个形状(N，)的numpy数组，其中包含用于minibatch的标签。
- reg: (float)正则化强度。
Returns: A tuple containing:
- loss as a single float
- 关于self.W的梯度;与W形状相同的数组
"""
loss = 0.0 # 初始化为0 float
dW = np.zeros(self.W.shape) # 与W形状相同的数组（初始化为0）
# 计算损失:
num_train = X_batch.shape[0] # 获取样本范围
scores = X_batch.dot(self.W)
correct_scores = scores[np.arange(num_train),y_batch]
margins = np.maximum(0, scores - correct_scores + 1)
loss += np.sum(margins) # 获取所有损失记录结果
loss /=num_train # 所有记录结果平均值
loss += reg * np.sum(self.W*self.W) # 正则化
# 计算梯度：
margins[margins > 0] = 1
row_num = -np.sum(margins, 1)
margins[np.arange(num_train), y] = row_num
dW += np.dot(X_batch.T, margins)/num_train +reg * self.W
然后使用下面的代码运行它：
# In the file linear_classifier.py, implement SGD in the function
# LinearClassifier.train() and then run it with the code below.
from cs231n.classifiers import LinearSVM
svm = LinearSVM()
tic = time.time()
loss_hist = svm.train(X_train, y_train, learning_rate=1e-7, reg=2.5e4,
num_iters=1500, verbose=True)
toc = time.time()
print('That took %fs' % (toc - tic))
输出：
iteration 0 / 1500: loss 799.173590
iteration 100 / 1500: loss 474.736085
iteration 200 / 1500: loss 288.689859
iteration 300 / 1500: loss 175.735395
iteration 400 / 1500: loss 107.535725
iteration 500 / 1500: loss 67.656875
iteration 600 / 1500: loss 42.015562
iteration 700 / 1500: loss 28.268911
iteration 800 / 1500: loss 18.718019
iteration 900 / 1500: loss 13.551437
iteration 1000 / 1500: loss 10.574625
iteration 1100 / 1500: loss 8.425946
iteration 1200 / 1500: loss 7.439375
iteration 1300 / 1500: loss 6.384190
iteration 1400 / 1500: loss 6.081119
That took 7.483288s
一个有用的调试策略是将损失绘制为迭代数的函数:
# A useful debugging strategy is to plot the loss as a function of
# iteration number:
plt.plot(loss_hist)
plt.xlabel('Iteration number')
plt.ylabel('Loss value')
plt.show()
写LinearSVM的预测功能。
def predict(self, X):
"""
Use the trained weights of this linear classifier to predict labels for
data points.
利用该线性分类器的训练权值预测数据点的标签
Inputs:
- X: 包含训练数据的形状(N, D)的numpy数组;每个维度D都有N个训练样本。
Returns:
- y_pred: x中数据的预测标签。y_pred是一个长度为N的一维数组，每个元素都是给出预测类的整数。
"""
y_pred = np.zeros(X.shape[0]) # 初始化
###########################################################################
# TODO: #
# Implement this method. Store the predicted labels in y_pred. #
# 将预测的标签存储在y_pred中 #
###########################################################################
# *****START OF YOUR CODE (DO NOT DELETE/MODIFY THIS LINE)*****
# x(N,D) W(D,C)
scores = X.dot(self.W) # (N , C)
y_pred = np.argmax(scores,axis=1) # 获取得到预测的类（1，N）
# *****END OF YOUR CODE (DO NOT DELETE/MODIFY THIS LINE)*****
return y_pred
写好预测功能后，评估性能训练集和验证集的性能
# Write the LinearSVM.predict function and evaluate the performance on both the
# training and validation set
y_train_pred = svm.predict(X_train)
print('training accuracy: %f' % (np.mean(y_train == y_train_pred), ))
y_val_pred = svm.predict(X_val)
print('validation accuracy: %f' % (np.mean(y_val == y_val_pred), ))
输出：
training accuracy: 0.382551
validation accuracy: 0.386000
使用验证集来调优超参数(正则化强度和学习率)。
您应该尝试不同的学习速率和正则化强度范围;如果您小心的话，您应该能够在验证集上获得大约0.39的分类精度。
# Use the validation set to tune hyperparameters (regularization strength and
# learning rate). You should experiment with different ranges for the learning
# rates and regularization strengths; if you are careful you should be able to
# get a classification accuracy of about 0.39 on the validation set.
#Note: you may see runtime/overflow warnings during hyper-parameter search.
# This may be caused by extreme values, and is not a bug.
# 注意:在超参数搜索期间，您可能会看到运行时/溢出警告。这可能是由极值引起的，而不是一个bug。
learning_rates = [1e-7, 5e-5]
regularization_strengths = [2.5e4, 5e4]
# results is dictionary mapping tuples of the form
# (learning_rate, regularization_strength) to tuples of the form
# (training_accuracy, validation_accuracy). The accuracy is simply the fraction
# of data points that are correctly classified.
# 结果是字典将表单元组(learning_rate、regularization_strength)映射为
# 表单元组(training_accuracy、validation_accuracy)。
# 精度只是正确分类的数据点的比例。
results = {}
best_val = -1 # The highest validation accuracy that we have seen so far.
best_svm = None # The LinearSVM object that achieved the highest validation rate.
################################################################################
# TODO: #
# Write code that chooses the best hyperparameters by tuning on the validation #
# set. For each combination of hyperparameters, train a linear SVM on the #
# training set, compute its accuracy on the training and validation sets, and #
# store these numbers in the results dictionary. In addition, store the best #
# validation accuracy in best_val and the LinearSVM object that achieves this #
# accuracy in best_svm. #
# 通过调整验证集来选择最佳超参数。对于每个超参数组合，在训练集上训练一个线性SVM，
# 在训练集和验证集上计算其精度，并将这些数字存储在结果字典中。
# 此外，将最佳验证精度存储在best_val中，而在best_svm中存储实现此精度的线性svm对象。
# Hint: You should use a small value for num_iters as you develop your #
# validation code so that the SVMs don't take much time to train; once you are #
# confident that your validation code works, you should rerun the validation #
# code with a larger value for num_iters. #
# 在开发验证代码时，应该为num_iter使用一个小值，这样SVMs就不会花费太多时间进行培训;
# 一旦您确信您的验证代码可以工作，您就应该为num_iter重新运行验证代码，并使用更大的值。
################################################################################
# *****START OF YOUR CODE (DO NOT DELETE/MODIFY THIS LINE)*****
for learning_rate in learning_rates:
for regularization_strength in regularization_strengths:
svm = LinearSVM() # 对于每个超参数组合，训练一个线性SVM
loss_history = svm.train(X_train, y_train,
learning_rate=learning_rate,
reg=regularization_strength,
num_iters=1500, verbose=True)
y_train_pred = svm.predict(X_train) # 这里只是对训练集进行预测而非精度
train_acc = np.mean(y_train == y_train_pred) # 在训练集上计算其精度
y_val_pred = svm.predict(X_val) # 这里只是对验证集进行预测而非精度
val_acc = np.mean(y_val == y_val_pred) # 在验证集上计算其精度
if val_acc > best_val: # 最佳验证精度存储在best_val中
best_val = val_acc
best_svm = svm # 同时获取实现此精度的线性svm对象
# results是将元组(learning_rate, regularization_strength)映射为
# (training_accuracy, validation_accuracy)的字典
results[(learning_rate, regularization_strength)] = [train_acc, val_acc]
# *****END OF YOUR CODE (DO NOT DELETE/MODIFY THIS LINE)*****
# Print out results.
for lr, reg in sorted(results):
train_accuracy, val_accuracy = results[(lr, reg)]
print('lr %e reg %e train accuracy: %f val accuracy: %f' % (
lr, reg, train_accuracy, val_accuracy))
print('best validation accuracy achieved during cross-validation: %f' % best_val)
这里result字典的赋值卡壳了，参照别人的写的，写法请引起注意，找错找了半天…
输出：
...
lr 1.000000e-07 reg 2.500000e+04 train accuracy: 0.380755 val accuracy: 0.377000
lr 1.000000e-07 reg 5.000000e+04 train accuracy: 0.365347 val accuracy: 0.366000
lr 5.000000e-05 reg 2.500000e+04 train accuracy: 0.169265 val accuracy: 0.181000
lr 5.000000e-05 reg 5.000000e+04 train accuracy: 0.055898 val accuracy: 0.048000
best validation accuracy achieved during cross-validation: 0.377000
可视化交叉验证结果：
# Visualize the cross-validation results
import math
x_scatter = [math.log10(x[0]) for x in results]
y_scatter = [math.log10(x[1]) for x in results]
# plot training accuracy
marker_size = 100
colors = [results[x][0] for x in results]
plt.subplot(2, 1, 1)
plt.scatter(x_scatter, y_scatter, marker_size, c=colors, cmap=plt.cm.coolwarm)
plt.colorbar()
plt.xlabel('log learning rate')
plt.ylabel('log regularization strength')
plt.title('CIFAR-10 training accuracy')
# plot validation accuracy
colors = [results[x][1] for x in results] # default size of markers is 20
plt.subplot(2, 1, 2)
plt.scatter(x_scatter, y_scatter, marker_size, c=colors, cmap=plt.cm.coolwarm)
plt.colorbar()
plt.xlabel('log learning rate')
plt.ylabel('log regularization strength')
plt.title('CIFAR-10 validation accuracy')
plt.show()
输出：
评估测试集上的最佳svm：
# Evaluate the best svm on test set
y_test_pred = best_svm.predict(X_test)
test_accuracy = np.mean(y_test == y_test_pred)
print('linear SVM on raw pixels final test set accuracy: %f' % test_accuracy)
输出：
linear SVM on raw pixels final test set accuracy: 0.370000
将每节课所学的权重形象化：
# Visualize the learned weights for each class.
# Depending on your choice of learning rate and regularization strength, these may
# or may not be nice to look at.
# 根据你对学习速度和正则化强度的选择，这些可能好看，也可能不好看。
w = best_svm.W[:-1,:] # strip out the bias
w = w.reshape(32, 32, 3, 10)
w_min, w_max = np.min(w), np.max(w)
classes = ['plane', 'car', 'bird', 'cat', 'deer', 'dog', 'frog', 'horse', 'ship', 'truck']
for i in range(10):
plt.subplot(2, 5, i + 1)
# Rescale the weights to be between 0 and 255
wimg = 255.0 * (w[:, :, :, i].squeeze() - w_min) / (w_max - w_min)
plt.imshow(wimg.astype('uint8'))
plt.axis('off')
plt.title(classes[i])
输出：
Inline Question2
描述您的可视化支持向量机权重是什么样子的，并提供一个简短的解释，为什么它们看起来是这样的。
Your Answier： 它们看起来像模糊信号，因为它学了数据集中的所有图片。（看别人的…）
扫一扫
2155
被折叠的 条评论
为什么被折叠?
到【灌水乐园】发言
