-
Notifications
You must be signed in to change notification settings - Fork 0
/
neural_network.py
173 lines (146 loc) · 7 KB
/
neural_network.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
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
# Name : Pranav Bhandari
# Student ID: 1001551132
# Date : 09/28/2020
import sys, random, math, numpy as np
def sigmoid(x):
try:
return 1 / (1 + math.exp(-x))
except:
print("Oops: Sigmoid Failed!")
return 0
def convertToOneHotVectors(data):
# The following code maps class lables to indices, which is then used to
# create one hot vectors for each training input
#
# The training inputs are then returned as data, and the target outputs
# for each training input is returned as one hot vectors
unique = np.unique(data)
indices = [i for i in range(len(unique))]
mapping = dict(zip(unique, indices))
reverse_mapping = dict(zip(indices, unique))
target_output = []
for i in range(len(data)):
vector = [0.0 for j in range(len(unique))]
vector[mapping.get(data[i])] = 1.0
target_output.append(vector)
return target_output, len(unique), reverse_mapping
def dataPreprocessing(filename):
# This function reads the file and normalizes the input
# It also creates one hot vectors for all the outputs
file = open(filename, "r")
data = []
output = []
max_value = sys.float_info.min
for row in file:
temp = row.split()
length = len(temp)
intermediate = []
for i in range(length-1):
intermediate.append(float(temp[i]))
if float(temp[i]) > max_value:
max_value = float(temp[i])
output.append(temp[length-1])
data.append(intermediate)
# Normalizing the attribute values with the MAXIMUM ABSOLUTE value over all attributes over all training objects for that dataset.
data = [[float(num/max_value) for num in row] for row in data]
return data, output
def training(training_file, layers, units_per_layer, rounds):
# Index starts from 0 and not 1 for both units and layers
# The training file is read and the associated target values are converted to one-hot vectors
training_input, output = dataPreprocessing(training_file)
training_output, num_classes, reverse_mapping = convertToOneHotVectors(output)
num_attributes = len(training_input[0])
# The following lines initialize the weights b and w to small random float values between -0.5 and 0.5
# The b and w values don't exist for the first layer and hence are not initialized
unitsInEachLayer = [units_per_layer for i in range(layers)]
unitsInEachLayer[0] = num_attributes
unitsInEachLayer[layers-1] = num_classes
b = [[] for i in range(layers)]
w = [[] for i in range(layers)]
for l in range(1, layers):
b[l] = [random.uniform(-0.05, 0.05) for i in range(unitsInEachLayer[l])]
w[l] = [[random.uniform(-0.05, 0.05) for j in range(unitsInEachLayer[l-1])] for i in range(unitsInEachLayer[l])]
learning_rate = 1.0
for r in range(rounds):
for n in range(len(training_input)):
z = [[] for i in range(layers)]
a = [[] for i in range(layers)]
z[0] = [0 for i in range(num_attributes)]
for i in range(num_attributes):
z[0][i] = training_input[n][i]
for l in range(1, layers):
a[l] = [0.0 for i in range(unitsInEachLayer[l])]
z[l] = [0.0 for i in range(unitsInEachLayer[l])]
for i in range(unitsInEachLayer[l]):
weighted_sum = 0.0
for j in range(unitsInEachLayer[l-1]):
weighted_sum += (w[l][i][j] * z[l-1][j])
a[l][i] = b[l][i] + weighted_sum
z[l][i] = sigmoid(a[l][i])
delta =[[] for i in range(layers)]
delta[layers-1] = [0 for i in range(num_classes)]
for i in range(num_classes):
delta[layers-1][i] = (z[layers-1][i] - training_output[n][i]) * z[layers-1][i] * (1.0-z[layers-1][i])
for l in range(layers-2, 0, -1):
delta[l] = [0 for i in range(unitsInEachLayer[l])]
for i in range(unitsInEachLayer[l]):
sum = 0.0
for k in range(unitsInEachLayer[l+1]):
sum += (delta[l+1][k] * w[l+1][k][i])
delta[l][i] = sum * z[l][i] * (1 - z[l][i])
for l in range(1, layers):
for i in range(unitsInEachLayer[l]):
b[l][i] -= (learning_rate * delta[l][i])
for j in range(unitsInEachLayer[l-1]):
w[l][i][j] -= (learning_rate * delta[l][i] * z[l-1][j])
learning_rate *= 0.98
return b, w, reverse_mapping, num_classes, unitsInEachLayer
def testing(test_file, layers, b, w, reverse_mapping, num_classes, unitsInEachLayer):
# Index starts from 0 and not 1 for both units and layers
test_input, test_output = dataPreprocessing(test_file)
num_attributes = len(test_input[0])
accuracy = 0.0
for n in range(len(test_input)):
z = [[] for i in range(layers)]
a = [[] for i in range(layers)]
z[0] = [0.0 for i in range(num_attributes)]
for i in range(num_attributes):
z[0][i] = test_input[n][i]
for l in range(1, layers):
a[l] = [0.0 for i in range(unitsInEachLayer[l])]
z[l] = [0.0 for i in range(unitsInEachLayer[l])]
for i in range(unitsInEachLayer[l]):
weighted_sum = 0.0
for j in range(unitsInEachLayer[l-1]):
weighted_sum += (w[l][i][j] * z[l-1][j])
a[l][i] = b[l][i] + weighted_sum
z[l][i] = sigmoid(a[l][i])
argmax = []
max_value = -1
for i in range(num_classes):
if z[layers-1][i] > max_value:
max_value = z[layers-1][i]
argmax.clear()
argmax.append(i)
elif z[layers-1][i] == max_value:
argmax.append(i)
predicted = [reverse_mapping.get(n) for n in argmax]
true = test_output[n]
actual_predicted = predicted[0]
if len(predicted)==1 and int(predicted[0]) == int(true):
curr_accuracy = 1.0
else:
try:
index = predicted.index(true)
actual_predicted = predicted[index]
curr_accuracy = float(1.0/len(predicted))
except ValueError:
curr_accuracy = 0.0
accuracy += curr_accuracy
print('ID={:5d}, predicted={:3d}, true={:3d}, accuracy={:4.2f}'.format(n+1, int(actual_predicted), int(true), curr_accuracy))
print('classification accuracy={:6.4f}'.format(accuracy/len(test_input)))
def neural_network(training_file, test_file, layers, units_per_layer, rounds):
b, w, reverse_mapping, num_classes, unitsInEachLayer = training(training_file, layers, units_per_layer, rounds)
testing(test_file, layers, b, w, reverse_mapping, num_classes, unitsInEachLayer)
if __name__ == '__main__':
neural_network(sys.argv[1], sys.argv[2], int(sys.argv[3]), int(sys.argv[4]), int(sys.argv[5]))