質問編集履歴
2
このような形でよろしいでしょうか？
title
CHANGED
|
File without changes
|
body
CHANGED
|
@@ -135,6 +135,10 @@
|
|
|
135
135
|
|
|
136
136
|
というコードに対し、
|
|
137
137
|
```
|
|
138
|
+
Using TensorFlow backend.
|
|
139
|
+
Traceback (most recent call last):
|
|
140
|
+
File "test2.py", line 102, in <module>
|
|
141
|
+
t_batch = train_T[batch_choice]
|
|
138
142
|
IndexError: index 44061 is out of bounds for axis 0 with size 10000
|
|
139
143
|
|
|
140
144
|
```
|
1
お願いします
title
CHANGED
|
File without changes
|
body
CHANGED
|
@@ -1,4 +1,102 @@
|
|
|
1
1
|
```
|
|
2
|
+
# -*- coding: utf-8 -*-
|
|
3
|
+
import keras
|
|
4
|
+
from keras.datasets import mnist
|
|
5
|
+
import matplotlib
|
|
6
|
+
from matplotlib import pyplot
|
|
7
|
+
import numpy as np
|
|
8
|
+
from sklearn import datasets
|
|
9
|
+
from sklearn.model_selection import cross_val_score as crv
|
|
10
|
+
|
|
11
|
+
def sigmoid(x):
|
|
12
|
+
return 1 / (1 + np.exp(-x))
|
|
13
|
+
|
|
14
|
+
def softmax(x):
|
|
15
|
+
expX = np.exp(x)
|
|
16
|
+
return expX / np.sum(expX)
|
|
17
|
+
|
|
18
|
+
def cross_entropy_error(y,t):
|
|
19
|
+
delta = 1e-7
|
|
20
|
+
batch_size = y.shape[0]
|
|
21
|
+
idx= np.arange(batch_size)
|
|
22
|
+
return -np.sum(np.log(y[idx,t]+delta)) / batch_size
|
|
23
|
+
|
|
24
|
+
def gradient(f,x):
|
|
25
|
+
if x.ndim == 1:
|
|
26
|
+
return gradient_sub(f,x)
|
|
27
|
+
else:
|
|
28
|
+
grad = np.zeros_like(x)
|
|
29
|
+
for index, xx in enumerate(x):
|
|
30
|
+
grad[index] = gradient_sub(f,xx)
|
|
31
|
+
return grad
|
|
32
|
+
|
|
33
|
+
def gradient_sub(f,x):
|
|
34
|
+
h = 1e-4
|
|
35
|
+
grad = np.zeros_like(x)
|
|
36
|
+
for i in range(x.size):
|
|
37
|
+
val = x[i]
|
|
38
|
+
x[i] = val + h
|
|
39
|
+
fx1 = f(x)
|
|
40
|
+
x[i] = val - h
|
|
41
|
+
fx2 = f(x)
|
|
42
|
+
grad[i] = (fx1 - fx2) / (2*h)
|
|
43
|
+
x[i] = val
|
|
44
|
+
return grad
|
|
45
|
+
|
|
46
|
+
|
|
47
|
+
class SampleNetwork:
|
|
48
|
+
def __init__(self):
|
|
49
|
+
input_size = 64
|
|
50
|
+
hidden_size = 50
|
|
51
|
+
output_size = 10
|
|
52
|
+
self.params = {}
|
|
53
|
+
self.params["w0"] = 0.01 * np.random.randn(input_size,hidden_size)
|
|
54
|
+
self.params["w1"] = 0.01 * np.random.randn(hidden_size,output_size)
|
|
55
|
+
self.params["b0"] = np.zeros(hidden_size)
|
|
56
|
+
self.params["b1"] = np.zeros(output_size)
|
|
57
|
+
|
|
58
|
+
self.learning_rate = 0.1
|
|
59
|
+
|
|
60
|
+
def predict(self,x):
|
|
61
|
+
a0 = np.dot(x,self.params["w0"]) + self.params["b0"]
|
|
62
|
+
z0 = sigmoid(a0)
|
|
63
|
+
a1 = np.dot(z0,self.params["w1"]) + self.params["b1"]
|
|
64
|
+
y = softmax(a1)
|
|
65
|
+
return y
|
|
66
|
+
|
|
67
|
+
def update_params(self,x,t):
|
|
68
|
+
loss_W = lambda W: self.loss(x,t)
|
|
69
|
+
for key in self.params.keys():
|
|
70
|
+
grad = gradient(loss_W, self.params[key])
|
|
71
|
+
self.params[key] -= self.learning_rate*grad
|
|
72
|
+
|
|
73
|
+
|
|
74
|
+
def loss(self,x,t):
|
|
75
|
+
y = self.predict(x)
|
|
76
|
+
return cross_entropy_error(y,t)
|
|
77
|
+
|
|
78
|
+
def accurary(self,x,t):
|
|
79
|
+
y = self.predict(x)
|
|
80
|
+
y = np.argmax(y,axis=1)
|
|
81
|
+
acc = np.sum(y==t) / float(x.shape[0])
|
|
82
|
+
return acc
|
|
83
|
+
|
|
84
|
+
digits = datasets.load_digits()
|
|
85
|
+
X = digits.data
|
|
86
|
+
T = digits.target
|
|
87
|
+
|
|
88
|
+
(train_X, test_X), (train_T,test_T ) = mnist.load_data()
|
|
89
|
+
|
|
90
|
+
network = SampleNetwork()
|
|
91
|
+
|
|
92
|
+
batch_size = 100
|
|
93
|
+
iter_num = 300
|
|
94
|
+
train_size = train_X.shape[0]
|
|
95
|
+
erpoch_size = max(train_size//batch_size,1)
|
|
96
|
+
loss_list = []
|
|
97
|
+
train_accurary_list = []
|
|
98
|
+
test_accurary_list = []
|
|
99
|
+
|
|
2
100
|
for index in range(iter_num):
|
|
3
101
|
batch_choice = np.random.choice(train_size,batch_size)
|
|
4
102
|
x_batch = train_X[batch_choice]
|
|
@@ -6,6 +104,33 @@
|
|
|
6
104
|
network.update_params(x_batch,t_batch)
|
|
7
105
|
loss = network.loss(x_batch,t_batch)
|
|
8
106
|
loss_list.append(loss)
|
|
107
|
+
if (index % erpoch_size == 0):
|
|
108
|
+
train_accurary = network.accurary(train_X,train_T)
|
|
109
|
+
test_accurary = network.accurary(test_X,test_T)
|
|
110
|
+
train_accurary_list.append(train_accurary)
|
|
111
|
+
test_accurary_list.append(test_accurary)
|
|
112
|
+
|
|
113
|
+
|
|
114
|
+
pyplot.figure(figsize=(10,7))
|
|
115
|
+
pyplot.subplot(2,2,1)
|
|
116
|
+
pyplot.plot(np.arange(len(loss_list)),loss_list)
|
|
117
|
+
pyplot.xlabel("iteration")
|
|
118
|
+
pyplot.title("Cross Entropy Error")
|
|
119
|
+
pyplot.subplot(2,2,2)
|
|
120
|
+
pyplot.plot(np.arange(0,len(train_accurary_list),1),train_accurary_list,"b")
|
|
121
|
+
pyplot.plot(np.arange(0,len(test_accurary_list),1),test_accurary_list,"ro")
|
|
122
|
+
pyplot.xlabel("iteration(epoch)")
|
|
123
|
+
pyplot.title("Accuary")
|
|
124
|
+
pyplot.legend(("train","test"),loc = "lowrer right")
|
|
125
|
+
pyplot.tight_layout()
|
|
126
|
+
|
|
127
|
+
|
|
128
|
+
|
|
129
|
+
|
|
130
|
+
|
|
131
|
+
|
|
132
|
+
|
|
133
|
+
|
|
9
134
|
```
|
|
10
135
|
|
|
11
136
|
というコードに対し、
|