Upload 9 files
Browse files- .gitattributes +6 -0
- all.py +223 -0
- cifar-10-batches-py/batches.meta +0 -0
- cifar-10-batches-py/data_batch_1 +3 -0
- cifar-10-batches-py/data_batch_2 +3 -0
- cifar-10-batches-py/data_batch_3 +3 -0
- cifar-10-batches-py/data_batch_4 +3 -0
- cifar-10-batches-py/data_batch_5 +3 -0
- cifar-10-batches-py/readme.html +1 -0
- cifar-10-batches-py/test_batch +3 -0
.gitattributes
CHANGED
@@ -32,3 +32,9 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
|
|
32 |
*.zip filter=lfs diff=lfs merge=lfs -text
|
33 |
*.zst filter=lfs diff=lfs merge=lfs -text
|
34 |
*tfevents* filter=lfs diff=lfs merge=lfs -text
|
|
|
|
|
|
|
|
|
|
|
|
|
|
32 |
*.zip filter=lfs diff=lfs merge=lfs -text
|
33 |
*.zst filter=lfs diff=lfs merge=lfs -text
|
34 |
*tfevents* filter=lfs diff=lfs merge=lfs -text
|
35 |
+
cifar-10-batches-py/data_batch_1 filter=lfs diff=lfs merge=lfs -text
|
36 |
+
cifar-10-batches-py/data_batch_2 filter=lfs diff=lfs merge=lfs -text
|
37 |
+
cifar-10-batches-py/data_batch_3 filter=lfs diff=lfs merge=lfs -text
|
38 |
+
cifar-10-batches-py/data_batch_4 filter=lfs diff=lfs merge=lfs -text
|
39 |
+
cifar-10-batches-py/data_batch_5 filter=lfs diff=lfs merge=lfs -text
|
40 |
+
cifar-10-batches-py/test_batch filter=lfs diff=lfs merge=lfs -text
|
all.py
ADDED
@@ -0,0 +1,223 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import os
|
2 |
+
from sklearn.preprocessing import OneHotEncoder # 独热编码
|
3 |
+
import numpy as np
|
4 |
+
import pickle as pk
|
5 |
+
|
6 |
+
|
7 |
+
# 载入数据
|
8 |
+
def load_batch(file): # 读取一个批次的数据
|
9 |
+
with open(file, 'rb') as f:
|
10 |
+
data_dict = pk.load(f, encoding='bytes')
|
11 |
+
images = data_dict[b'data']
|
12 |
+
labels = data_dict[b'labels']
|
13 |
+
images = images.reshape(10000, 3072)
|
14 |
+
labels = np.array(labels)
|
15 |
+
return (images / 255), labels
|
16 |
+
|
17 |
+
|
18 |
+
def load_data(data_dir):
|
19 |
+
images_train = []
|
20 |
+
labels_train = []
|
21 |
+
for i in range(5):
|
22 |
+
file = os.path.join(data_dir, 'data_batch_%d' % (i + 1))
|
23 |
+
print('加载文件:', file)
|
24 |
+
# 按批次读取训练集数据并拼接到图像和标签列表后,直到读入所有批次数据
|
25 |
+
images_batch, labels_batch = load_batch(file)
|
26 |
+
images_train.append(images_batch)
|
27 |
+
labels_train.append(labels_batch)
|
28 |
+
# 将多个批次的数组统一为一个数组
|
29 |
+
x_train = np.concatenate(images_train)
|
30 |
+
t_train = np.concatenate(labels_train)
|
31 |
+
del images_batch, labels_batch
|
32 |
+
|
33 |
+
# 加载测试集图像和标签
|
34 |
+
x_test, t_test = load_batch(os.path.join(data_dir, 'test_batch'))
|
35 |
+
return x_train, t_train, x_test, t_test
|
36 |
+
|
37 |
+
|
38 |
+
def sigmoid(x):
|
39 |
+
return 1 / (1 + np.exp(-x))
|
40 |
+
|
41 |
+
|
42 |
+
def sigmoid_grad(x):
|
43 |
+
return (1.0 - sigmoid(x)) * sigmoid(x)
|
44 |
+
|
45 |
+
|
46 |
+
def softmax(x):
|
47 |
+
if x.ndim == 2:
|
48 |
+
x = x.T
|
49 |
+
x = x - np.max(x, axis=0)
|
50 |
+
y = np.exp(x) / np.sum(np.exp(x), axis=0)
|
51 |
+
return y.T
|
52 |
+
|
53 |
+
x = x - np.max(x) # 溢出对策
|
54 |
+
return np.exp(x) / np.sum(np.exp(x))
|
55 |
+
|
56 |
+
|
57 |
+
class neuralNetwork:
|
58 |
+
|
59 |
+
def __init__(self, numNeuronLayers, numNeurons_perLayer, learningRate):
|
60 |
+
self.numNeurons_perLayer = numNeurons_perLayer
|
61 |
+
self.numNeuronLayers = numNeuronLayers
|
62 |
+
self.learningRate = learningRate
|
63 |
+
self.weight = []
|
64 |
+
self.bias = []
|
65 |
+
for i in range(numNeuronLayers):
|
66 |
+
self.weight.append(
|
67 |
+
learningRate * np.random.randn(self.numNeurons_perLayer[i], self.numNeurons_perLayer[i + 1]))
|
68 |
+
self.bias.append(np.zeros(self.numNeurons_perLayer[i + 1]))
|
69 |
+
|
70 |
+
def predict(self, x):
|
71 |
+
z = x
|
72 |
+
# 走一遍前向传播得到输出
|
73 |
+
for i in range(self.numNeuronLayers - 1):
|
74 |
+
a = np.dot(z, self.weight[i]) + self.bias[i]
|
75 |
+
z = sigmoid(a)
|
76 |
+
an = np.dot(z, self.weight[self.numNeuronLayers - 1]) + self.bias[self.numNeuronLayers - 1]
|
77 |
+
y = softmax(an)
|
78 |
+
return y
|
79 |
+
|
80 |
+
def gradient(self, x, t):
|
81 |
+
z = []
|
82 |
+
a = []
|
83 |
+
z.append(x)
|
84 |
+
# 走一遍前向传播得到输出
|
85 |
+
for i in range(self.numNeuronLayers):
|
86 |
+
a.append(np.dot(z[i], self.weight[i]) + self.bias[i])
|
87 |
+
z.append(sigmoid(a[i]))
|
88 |
+
y = softmax(a[self.numNeuronLayers - 1])
|
89 |
+
num = x.shape[0]
|
90 |
+
dy = (y - t) / num
|
91 |
+
dz = []
|
92 |
+
da = []
|
93 |
+
dz.append(dy)
|
94 |
+
for i in range(self.numNeuronLayers - 1):
|
95 |
+
da.append(np.dot(dz[i], self.weight[self.numNeuronLayers - i - 1].T))
|
96 |
+
dz.append(sigmoid_grad(a[self.numNeuronLayers - i - 2]) * da[i])
|
97 |
+
|
98 |
+
for i in range(self.numNeuronLayers):
|
99 |
+
self.weight[i] -= self.learningRate * np.dot(z[i].T, dz[self.numNeuronLayers - i - 1])
|
100 |
+
self.bias[i] -= self.learningRate * np.sum(dz[self.numNeuronLayers - i - 1], axis=0)
|
101 |
+
|
102 |
+
def loss(self, x, t):
|
103 |
+
y = self.predict(x)
|
104 |
+
t = t.argmax(axis=1)
|
105 |
+
num = y.shape[0]
|
106 |
+
s = y[np.arange(num), t]
|
107 |
+
return -np.sum(np.log(s)) / num
|
108 |
+
|
109 |
+
def accuracy(self, x, t):
|
110 |
+
y = self.predict(x)
|
111 |
+
p = np.argmax(y, axis=1)
|
112 |
+
q = np.argmax(t, axis=1)
|
113 |
+
acc = np.sum(p == q) / len(y)
|
114 |
+
return acc
|
115 |
+
|
116 |
+
|
117 |
+
def kNN(x_train, x_test, t_train, k):
|
118 |
+
px = list()
|
119 |
+
for i in range(len(x_test)):
|
120 |
+
px.append([])
|
121 |
+
for j in range(10):
|
122 |
+
px[i].append(0)
|
123 |
+
for i in range(len(x_test)):
|
124 |
+
dis = getODistance(x_test[i], x_train)
|
125 |
+
index = np.argsort(dis)
|
126 |
+
count = list()
|
127 |
+
r = np.sort(dis)[k - 1]
|
128 |
+
for j in range(len(t_train[0])):
|
129 |
+
count.append(0)
|
130 |
+
for j in range(k):
|
131 |
+
for w in range(10):
|
132 |
+
if t_train[index[j]][w] == 1:
|
133 |
+
count[w] = count[w] + 1
|
134 |
+
for j in range(10):
|
135 |
+
px[i][j] = count[j]
|
136 |
+
return px
|
137 |
+
|
138 |
+
|
139 |
+
def getODistance(sample, train):
|
140 |
+
a = np.tile(sample, [1000, 1])
|
141 |
+
a = a - train
|
142 |
+
a = np.square(a)
|
143 |
+
a = a.sum(axis=1)
|
144 |
+
dis = np.sqrt(a)
|
145 |
+
dis = dis.T
|
146 |
+
dis = dis.tolist()
|
147 |
+
return dis[0]
|
148 |
+
|
149 |
+
|
150 |
+
def runNetwork():
|
151 |
+
numNeuronLayers = 3
|
152 |
+
numNeurons_perLayer = [3072, 50, 20, 10]
|
153 |
+
learningRate = 0.05
|
154 |
+
epoch = 50000
|
155 |
+
batch_size = 100
|
156 |
+
train_size = x_train.shape[0] # 50000
|
157 |
+
|
158 |
+
net = neuralNetwork(numNeuronLayers, numNeurons_perLayer, learningRate)
|
159 |
+
for i in range(epoch):
|
160 |
+
batch_mask = np.random.choice(train_size, batch_size) # 从0到50000 随机选100个数
|
161 |
+
x_batch = x_train[batch_mask]
|
162 |
+
t_batch = t_train[batch_mask]
|
163 |
+
net.gradient(x_batch, t_batch)
|
164 |
+
y = net.predict(x_test[0:1000, 0:3072])
|
165 |
+
p = np.argmax(y, axis=1)
|
166 |
+
q = np.argmax(t_test[0:1000, 0:3072], axis=1)
|
167 |
+
acc = np.sum(p == q) / len(y)
|
168 |
+
print("神经网络正确率为:", acc)
|
169 |
+
return p
|
170 |
+
|
171 |
+
|
172 |
+
def runKnn(x_train, x_test):
|
173 |
+
x_train = np.mat(x_train)
|
174 |
+
x_test = np.mat(x_test)
|
175 |
+
px = kNN(x_train[0:1000, 0:3072], x_test[0:1000, 0:3072], t_train[0:1000, 0:10], 7)
|
176 |
+
p = np.argmax(px, axis=1)
|
177 |
+
q = np.argmax(t_test[0:1000, 0:3072], axis=1)
|
178 |
+
acc = np.sum(p == q) / 1000
|
179 |
+
print("knn正确率为:", acc)
|
180 |
+
return p
|
181 |
+
|
182 |
+
|
183 |
+
from sklearn import svm
|
184 |
+
|
185 |
+
|
186 |
+
def runSvm():
|
187 |
+
clf = svm.SVC(probability=True)
|
188 |
+
t = np.argmax(t_train[0:1000, 0:3072], axis=1)
|
189 |
+
clf.fit(x_train[0:1000, 0:3072], t)
|
190 |
+
p = clf.predict(x_test[0:1000, 0:3072])
|
191 |
+
q = np.argmax(t_test[0:1000, 0:3072], axis=1)
|
192 |
+
acc = np.sum(p == q) / 1000
|
193 |
+
print("svm正确率为:", acc)
|
194 |
+
return p
|
195 |
+
|
196 |
+
|
197 |
+
data_dir = 'cifar-10-batches-py'
|
198 |
+
x_train, t_train, x_test, t_test = load_data(data_dir)
|
199 |
+
encoder = OneHotEncoder(sparse=False)
|
200 |
+
one_format = [[0], [1], [2], [3], [4], [5], [6], [7], [8], [9]]
|
201 |
+
encoder.fit(one_format)
|
202 |
+
t_train = t_train.reshape(-1, 1) # 数组化为一维包含一个元素的二维数组,-1代表二维的数量自适应
|
203 |
+
t_train = encoder.transform(t_train)
|
204 |
+
t_test = t_test.reshape(-1, 1)
|
205 |
+
t_test = encoder.transform(t_test)
|
206 |
+
|
207 |
+
p1 = runNetwork()
|
208 |
+
p2 = runSvm()
|
209 |
+
p3 = runKnn(x_train, x_test)
|
210 |
+
|
211 |
+
p1 = p1.reshape(-1, 1) # 数组化为一维包含一个元素的二维数组,-1代表二维的数量自适应
|
212 |
+
p1 = encoder.transform(p1)
|
213 |
+
p2 = p2.reshape(-1, 1)
|
214 |
+
p2 = encoder.transform(p2)
|
215 |
+
p3 = p3.reshape(-1, 1)
|
216 |
+
p3 = encoder.transform(p3)
|
217 |
+
|
218 |
+
vote = p1+p2+p3
|
219 |
+
p = np.argmax(vote, axis=1)
|
220 |
+
q = np.argmax(t_test[0:1000, 0:3072], axis=1)
|
221 |
+
acc = np.sum(p == q) / 1000
|
222 |
+
print("最终正确率为", acc)
|
223 |
+
|
cifar-10-batches-py/batches.meta
ADDED
Binary file (158 Bytes). View file
|
|
cifar-10-batches-py/data_batch_1
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:54636561a3ce25bd3e19253c6b0d8538147b0ae398331ac4a2d86c6d987368cd
|
3 |
+
size 31035704
|
cifar-10-batches-py/data_batch_2
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:766b2cef9fbc745cf056b3152224f7cf77163b330ea9a15f9392beb8b89bc5a8
|
3 |
+
size 31035320
|
cifar-10-batches-py/data_batch_3
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:0f00d98ebfb30b3ec0ad19f9756dc2630b89003e10525f5e148445e82aa6a1f9
|
3 |
+
size 31035999
|
cifar-10-batches-py/data_batch_4
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:3f7bb240661948b8f4d53e36ec720d8306f5668bd0071dcb4e6c947f78e9682b
|
3 |
+
size 31035696
|
cifar-10-batches-py/data_batch_5
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:d91802434d8376bbaeeadf58a737e3a1b12ac839077e931237e0dcd43adcb154
|
3 |
+
size 31035623
|
cifar-10-batches-py/readme.html
ADDED
@@ -0,0 +1 @@
|
|
|
|
|
1 |
+
<meta HTTP-EQUIV="REFRESH" content="0; url=http://www.cs.toronto.edu/~kriz/cifar.html">
|
cifar-10-batches-py/test_batch
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:f53d8d457504f7cff4ea9e021afcf0e0ad8e24a91f3fc42091b8adef61157831
|
3 |
+
size 31035526
|