-
Notifications
You must be signed in to change notification settings - Fork 0
/
Neural_Network_Relu.py
232 lines (154 loc) · 7.71 KB
/
Neural_Network_Relu.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
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
# -*- coding: utf-8 -*-
"""
Created on Thu Nov 8 17:44:45 2018
@author: AARUSHI
"""
"""Ref :- https://hackernoon.com/dl02-writing-a-neural-network-from-scratch-code-b32f4877c257
https://github.com/sar-gupta/neural-network-from-scratch/blob/master/neuralnetwork.py """
import numpy as np
class layer:
def __init__(self, num_nodes, layer_num,batch_size):
self.num_nodes = num_nodes
self.activation_function = "Sigmoid"
self.is_last = False
self.weight = None
self.bias = None
self.changed_weight = None
self.changed_bias = None
self.input_activation = None
self.input =None
if(layer_num==len(num_nodes)-1):
self.is_last = True
self.activation_function = "Softmax"
else:
self.weight = np.random.normal(0, 0.1, size=(num_nodes[layer_num], num_nodes[layer_num+1]))
self.bias = np.random.normal(0, 0.1, size=(1, num_nodes[layer_num+1]))
self.changed_weight = np.zeros((num_nodes[layer_num], num_nodes[layer_num+1]),dtype = float)
self.changed_bias = np.zeros((1, num_nodes[layer_num+1]),dtype = float)
class Neural_Network(object):
def __init__(self, nlayer, num_node, batch_size):
self.nlayer = nlayer
self.num_node= num_node
self.layers = []
self.batch_size = batch_size
for i in range(nlayer):
layer_i = layer(num_node, i,self.batch_size)
self.layers.append(layer_i)
def softmax(self, layer):
num = np.exp(layer)
den = np.sum(num)
return num/den
def relu(self,x):
t = np.copy(x)
x = t.shape[0]
y = t.shape[1]
for i in range(x):
for j in range(y):
if(t[i][j]<0):
t[i][j]=0.01*(t[i][j])
return t
def derivative_relu(self,x):
t = np.copy(x)
x = t.shape[0]
y = t.shape[1]
for i in range(x):
for j in range(y):
if(t[i][j]<0):
t[i][j]=0.01
else:
t[i][j]=1
return t
def cross_entropy_derivative(self,p,y):
return (p-y)
def train(self,batch_size, input, labels,x_test,y_test, num_epochs,learning_rate, filename):
self.learning_rate = learning_rate
num_batches = int(len(input)/batch_size)
print("Started")
for i in range(num_epochs):
for j in range(num_batches):
for k in range(j*batch_size , j*batch_size+batch_size):
output = (np.zeros((self.num_node[self.nlayer-1], 1)))
output[labels[k]]=1
output = output.T
#Forward Pass
self.forward_pass(input[k])
##ORIGINAL
###self.error -= (np.sum((labels[j*batch_size : j*batch_size+batch_size] * np.log(self.layers[self.nlayer-1].input_activation ))))
#Backward Pass
self.back_pass(output)
##Gradient Descent
self.grad_descent()
print("Epoch "+str(i)+" done")
self.check_accuracy('model.pkl',x_test,y_test)
print("Training Done")
def forward_pass(self,input):
self.layers[0].input_activation = np.reshape(input,(1,input.shape[0]))
self.layers[0].input = np.reshape(input,(1,input.shape[0]))
for i in range(self.nlayer-1):
first_term = np.dot(self.layers[i].input_activation,self.layers[i].weight)
second_term = self.layers[i].bias
curr_temp = np.add(first_term ,second_term)
self.layers[i+1].input = curr_temp
if(self.layers[i+1].activation_function=="Sigmoid"):
self.layers[i+1].input_activation = self.relu(curr_temp)
else:
self.layers[i+1].input_activation = self.softmax(curr_temp)
def back_pass(self, labels):
# if self.cost_function == "cross_entropy" and self.layers[self.num_layers-1].activation_function == "softmax":
layer = self.layers[self.nlayer-1]
prev_layer = self.layers[self.nlayer-2]
delta = self.cross_entropy_derivative(layer.input_activation,labels)
prev_layer.changed_bias += delta
prev_layer.changed_weight += np.dot(prev_layer.input_activation.T, delta)
for i in range(self.nlayer-2,0,-1):
layer = self.layers[i]
prev_layer = self.layers[i-1]
#print("jjojo")
#print(delta.shape)
#print(np.dot(delta,layer.weight.T).shape)
#print(self.sigmoid(layer.input_activation,True).T.shape)
delta = np.multiply(np.dot(delta,layer.weight.T), self.derivative_relu(layer.input))
#print(delta.shape)
#print(prev_layer.changed_bias.shape)
prev_layer.changed_bias += delta
#print(delta.shape)
#print(prev_layer.input_activation.T.shape)
#print("byeeeeeeeeeeeeeeeeeeeee")
prev_layer.changed_weight += np.dot(prev_layer.input_activation.T, delta)
def grad_descent(self):
for i in range(0,self.nlayer-1):
layer = self.layers[i]
layer.weight = layer.weight - (self.learning_rate/self.batch_size)*layer.changed_weight
layer.bias = layer.bias - (self.learning_rate/self.batch_size)*layer.changed_bias
layer.changed_weight = np.zeros((self.num_node[i], self.num_node[i+1]),dtype = float)
layer.changed_bias = np.zeros((1, self.num_node[i+1]),dtype = float)
def forward_pass_testing(self,input):
self.layers[0].input_activation = input
for i in range(self.nlayer-1):
first_term = np.dot(self.layers[i].input_activation,self.layers[i].weight)
second_term = self.layers[i].bias
curr_temp = np.add(first_term ,second_term)
if(self.layers[i+1].activation_function=="Sigmoid"):
self.layers[i+1].input_activation = self.relu(curr_temp)
else:
self.layers[i+1].input_activation = self.softmax(curr_temp)
def predict(self, filename, input):
# self.batch_size = len(input)
self.forward_pass_testing(input)
a = self.layers[self.nlayer-1].input_activation
predictions = []
for i in range(len(a)):
if(a[i][0]>a[i][1]):
predictions.append(0)
else:
predictions.append(1)
return np.asarray(predictions),a
def check_accuracy(self, filename, inputs, labels):
print("Testing Started")
a, array = self.predict(filename,inputs)
total=len(a)
correct=0
for i in range(len(a)):
if (a[i] == labels[i]):
correct += 1
print("Accuracy: ", correct*100/total)