-
Notifications
You must be signed in to change notification settings - Fork 0
/
BPNetwork.py
151 lines (131 loc) · 5.86 KB
/
BPNetwork.py
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
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
#!python3
from utility import *
import numpy as np
class FFDeco:
def __init__(self, v, w, gamma, theta):
self.v = v
self.w = w
self.gamma = gamma
self.theta = theta
def __call__(self, x):
return feed_forward(x, self.v, self.w, self.gamma, self.theta)
def sigmoid(x):
return 1 / (1 + np.exp(-x))
def feed_forward(x, v, w, gamma, theta):
"""
feed forward function in BP
:param x: input vector
:param v: 1st transfer matrix
:param w: 2nd transfer matrix
:param gamma: 1st threshold
:param theta: 2nd threshold
:return: b, y_hat
"""
alpha = np.dot(v, x)
b = sigmoid(alpha - gamma)
beta = np.dot(w, b)
y_hat = sigmoid(beta - theta)
return b, y_hat
def standard_back_propagation(eta, x, y, b, y_hat, w):
"""
Standard BP routine
:param eta: learning rate
:param x: input vector
:param y: training result vector
:param b: output of hidden layer
:param y_hat: estimated result (output of output layer)
:param w: transform matrix from hidden layer to output layer
:return: delta_w, delta_theta, delta_v, delta_gamma
"""
g = y_hat * (1 - y_hat) * (y - y_hat)
e = b * (1 - b) * np.dot(np.transpose(w), g)
delta_w = np.array(eta * np.matrix(g).T * np.matrix(b))
delta_theta = -eta * g
delta_v = np.array(eta * np.matrix(e).T * np.matrix(x))
delta_gamma = -eta * e
return delta_w, delta_theta, delta_v, delta_gamma
if __name__ == '__main__':
import argparse
from tqdm import tqdm
from multiprocessing import Pool
parser = argparse.ArgumentParser(description='Back propagation neural networks algorithm')
parser.add_argument('-e', '--eta', dest='eta', nargs='?', default=0.5, type=float, help='eta (learning rate) value')
parser.add_argument('--hidden', dest='hidden', nargs='?', default=250, type=int,
help='size of hidden layer')
parser.add_argument('-i', '--iter', dest='iter', nargs='?', default=100, type=int,
help='maximum iteration before stop training')
parser.add_argument('--err', '--error', dest='error', nargs='?', default=10, type=int,
help='desired training error rate')
parser.add_argument('-t', '--train', dest='train', nargs='?', default='', type=str, help='train result filename')
parser.add_argument('-r', '--result', dest='result', nargs='?', default='', type=str, help='test result filename')
parser.add_argument('-j', '--jobs', dest='jobs', nargs='?', default=1, type=int, help='maximum parallel jobs allowed')
args = parser.parse_args()
timer = UniversalTimer()
print("Loading training set...")
labelSpace = list(range(10))
trainingData, trainingLabel = unpack_mnist('Data/train-images.idx3-ubyte', 'Data/train-labels.idx1-ubyte')
#trainingData = trainingData[0:2000]
#trainingLabel = trainingLabel[0:2000]
t_mean = np.mean(trainingData)
t_std = np.std(trainingData)
trainingData -= t_mean
trainingData /= t_std
labelSpace = np.array(labelSpace)
# Standard BP
v = np.random.rand(args.hidden, trainingData.shape[1])
w = np.random.rand(labelSpace.shape[0], args.hidden)
gamma = np.zeros(args.hidden)
theta = np.zeros(labelSpace.shape[0])
y = np.zeros((trainingData.shape[0], labelSpace.shape[0]))
y[np.array(range(trainingLabel.shape[0])), trainingLabel] = 1
round = 0
training_error_rate = 100
while training_error_rate > args.error and round < args.iter:
# train training set once
train_error = 0
for i, x in enumerate(tqdm(trainingData, total=trainingData.shape[0], ascii=True, ncols=75, leave=False)):
b, y_hat = feed_forward(x, v, w, gamma, theta)
if np.argmax(y_hat) != trainingLabel[i]:
train_error += 1
delta_w, delta_theta, delta_v, delta_gamma = standard_back_propagation(args.eta, x, y[i], b, y_hat, w)
w += delta_w
theta += delta_theta
v += delta_v
gamma += delta_gamma
'''
# if some of training set were assigned to test the training result
result = list(map(lambda x: feed_forward(x, v, w, gamma, theta), trainingData))
train_test_error = 0
for i, r in enumerate(result):
if np.argmax(r[1]) != trainingLabel[i]:
train_test_error += 1
'''
round += 1
training_error_rate = train_error / trainingData.shape[0] * 100
tqdm.write("round %d, training error rate %.2f%%" % (round, training_error_rate))
if args.train:
with open(args.train, 'a') as test_result_file:
# TODO: if training set is split into two parts, update train test error here
test_result_file.write("%d, %.2f, %.2f\n" % (round, training_error_rate, 0))
print("Loading testing set...")
testingData, testingLabel = unpack_mnist('Data/t10k-images.idx3-ubyte', 'Data/t10k-labels.idx1-ubyte')
testingData -= t_mean
testingData /= t_std
result = list(map(lambda x: feed_forward(x, v, w, gamma, theta), testingData))
test_error = 0
for i, r in enumerate(result):
if np.argmax(r[1]) != testingLabel[i]:
test_error += 1
print_string = "Classification completed.\n" \
"Runtime = %.2f minutes\n" \
"Training samples = %d\n" \
"Training error rate = %2.2f%%\n" \
"Testing samples = %d\n" \
"Errors = %d\n" \
"Error rate = %2.2f%%\n"\
% (timer.current_time() / 60, trainingLabel.shape[0], training_error_rate, testingLabel.shape[0],
test_error, test_error / testingLabel.shape[0] * 100)
if args.result != '':
with open(args.result, 'w') as test_result_file:
test_result_file.write(print_string)
print('\n', print_string)