-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathneuralnetworknumpyOOP.py
181 lines (142 loc) · 6.27 KB
/
neuralnetworknumpyOOP.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
import numpy as np
import pandas as pd
import matplotlib.pyplot as plt
from tensorflow.keras.datasets import mnist
class NeuralNetwork(object):
"""
Arguments:
input - input should be (no. of instances, columns/features)
output - output should be (no. of instances, columns/features)
hiddensize - must be a list
- specify the size and the number of hidden layers
- default: None, which means no hidden layer.
example:
hiddenSize = [100, 50, 30] means:
1st hidden layer has 100 neurons,
2nd hidden layer has 50 neurons and
3rd hidden layer has 30 neurons.
Number of hidden layers is 3.
"""
def __init__(self, input_, output_, hiddenSize=None):
super(NeuralNetwork, self).__init__()
self.inputSize = input_.shape[1]
self.outputSize = output_.shape[1]
self.hiddenLen = len(hiddenSize)
self.weights = []
self.biases = []
if hiddenSize is None:
self.weights.append(np.random.randn(self.inputSize, self.outputSize))
self.biases.append(np.zeros((1, self.outputSize)))
else:
for i in range(len(hiddenSize)): # [30] [50, 50, 50, 25, 12] [2]
if i == 0: # if input X first hidden layer
self.weights.append(np.random.randn(self.inputSize, hiddenSize[i]))
self.biases.append(np.zeros((1, hiddenSize[i])))
elif i == len(hiddenSize) - 1: #if last hidden layer X output layer
self.weights.append(np.random.randn(hiddenSize[i-1], hiddenSize[i]))
self.biases.append(np.zeros((1, hiddenSize[i])))
self.weights.append(np.random.randn(hiddenSize[-1], self.outputSize))
self.biases.append(np.zeros((1, self.outputSize)))
break
else: #if nth hidden layer X nth hidden layer (n is not equal to first or last)
self.weights.append(np.random.randn(hiddenSize[i-1], hiddenSize[i]))
self.biases.append(np.zeros((1, hiddenSize[i])))
self.pre_activations = [0 for i in range(len(self.weights))]
self.activations = [0 for i in range(len(self.biases))]
self.nabla_w = [0 for i in range(len(self.weights))]
self.nabla_b = [0 for i in range(len(self.biases))]
def checkParameters(self):
print("Number of Weights: {}".format(len(self.weights)))
print("Number of Biases: {}".format(len(self.biases)))
print("Number of Pre_activations: {}".format(len(self.pre_activations)))
print("Number of Activations: {}".format(len(self.activations)))
for x, w in enumerate(self.weights):
print("Weight {} shape : {}".format(x+1, w.shape))
for x, b in enumerate(self.biases):
print("Bias {} shape : {}".format(x+1, b.shape))
for x, p in enumerate(self.pre_activations):
print("Pre_activation {} shape : {}".format(x+1, p.shape))
for x, a in enumerate(self.activations):
print("Activation {} shape : {}".format(x+1, a.shape))
def fit(self, input_, output_, epochs=10, lr=0.0001):
for epoch in range(epochs):
print("Epoch {}".format(epoch+1))
self.feedforward(input_)
self.backpropagation(input_,output_)
self.update_params(lr)
self.validate_score(output_)
self.accuracy_metrics(output_)
def feedforward(self, input_):
activations = input_
for i in range(len(self.weights)): # or self.biases
self.pre_activations[i] = np.add(activations.dot(self.weights[i]), self.biases[i])
activations = self.sigmoid(self.pre_activations[i])
self.activations[i] = activations
def backpropagation(self, input_, output_):
lossFunction = self.diff_cost_function(output_, self.activations[-1])*self.diff_sigmoid(self.pre_activations[-1])
loss = lossFunction
#print(type(lossFunction)) # for debugging
for i in range(len(self.weights)-1,-1,-1): # backward range: 4, 3, 2 ,1 ,0
if i==0:
self.nabla_w[i] = np.dot(input_.T, loss )
self.nabla_b[i] = np.mean(loss, axis=0).reshape(1,-1)
break
#print(type(self.activations[i-1])) # for debugging
self.nabla_w[i] = np.dot(self.activations[i-1].T , loss)
self.nabla_b[i] = np.mean(loss, axis=0).reshape(1,-1)
loss = (loss.dot(self.weights[i].T))*(self.diff_sigmoid(self.pre_activations[i-1]))
def update_params(self, lr):
for i in range(len(self.weights)):
self.weights[i] = self.weights[i] - lr*self.nabla_w[i]
self.biases[i] = self.biases[i] - lr*self.nabla_b[i]
def predict(self, input_, output_):
pre_activations_pred = [0 for i in range(len(self.biases))]
activations_pred = [0 for i in range(len(self.biases))]
activations = input_
for i in range(len(self.weights)): # or self.biases
pre_activations_pred[i] = np.add(activations.dot(self.weights[i]), self.biases[i])
activations = self.sigmoid(pre_activations_pred[i])
activations_pred[i] = activations
return activations_pred[-1]
def validate_score(self, output_):
mse = np.mean(np.sum(self.cost_function(output_, self.activations[-1]), axis=1))
print("Mean Squared Error: {}".format(mse))
def accuracy_metrics(self, output_, threshold=0.7):
self.correct = 0
self.wrong = 0
for i in range(len(output_)):
y = np.array([1 if max(self.activations[-1][i])==j else 0 for j in self.activations[-1][i]])
print(y.shape)
print(output_.shape)
y_num = np.where(y==1)[0][0]
true_y = np.where(output_[i]==1)[0][0]
if y_num == true_y:
self.correct += 1
else:
self.wrong += 1
print(self.correct)
print(self.wrong)
print("Accuracy: {:2f} %".format(self.correct/len(output_)*100))
def sigmoid(self, z):
return 1/(1+np.exp(z))
def diff_sigmoid(self, z):
return self.sigmoid(z)*(1-self.sigmoid(z))
def cost_function(self, y_train, output_):
return 0.5*np.square(y_train - output_)
def diff_cost_function(self, y_train, output_):
return (y_train - output_)
def extract_data():
(X_train, y_train),(X_test, y_test) = mnist.load_data()
X_train = X_train.reshape(60000, -1) / 255
X_test = X_test.reshape(10000, -1) / 255
y_train = pd.get_dummies(y_train).values
y_test = pd.get_dummies(y_test).values
return X_train, y_train, X_test, y_test
if __name__ == '__main__':
X_train, y_train, X_test, y_test = extract_data()
h_size = X_train.shape[1]
nn = NeuralNetwork(X_train, y_train, hiddenSize=[h_size, int(h_size/2), int(h_size/2), int(h_size/3), int(h_size/4)])
nn.fit(X_train, y_train, epochs=3)
y_pred = nn.predict(X_test, y_test)
print(y_pred.shape)
#nn.checkParameters()