linear_gradient_algorithm/neural_network.py at main · democheng/linear_gradient_algorithm · GitHub

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
import numpy as np
import random

class NeuralNetwork(object):
    def __init__(self, node_list, \
                activate_func, jacobian_of_activate_func, \
                cost_func, jacobian_of_cost_func) -> None:
        self.m_node_list = node_list
        # self.m_weight_list = [np.random.randn(num_from,num_to) for num_from,num_to in zip(node_list[:-1], node_list[1:])]
        # self.m_bias_list = [np.random.randn(num_to,1) for num_to in node_list[1:]]
        self.m_weight_list = [np.random.uniform(-2,0,(num_from,num_to))+1 for num_from,num_to in zip(node_list[:-1], node_list[1:])]
        self.m_bias_list = [np.random.uniform(-2,0,(num_to,1))+1 for num_to in node_list[1:]]
        # print('self.m_weight_list:', self.m_weight_list)
        # print('self.m_bias_list:', self.m_bias_list)
        self.m_activate_func = activate_func
        self.m_jacobian_of_activate_func = jacobian_of_activate_func
        self.m_cost_func = cost_func
        self.m_jacobian_of_cost_func = jacobian_of_cost_func

    def resetParameters(self):
        # for idx in range(len(self.m_weight_list)):
        #     self.m_weight_list[idx] = np.random.randn(self.m_weight_list[idx].shape[0],self.m_weight_list[idx].shape[1])
        # for idx in range(len(self.m_bias_list)):
        #     self.m_bias_list[idx] = np.random.randn(self.m_bias_list[idx].shape[0],self.m_bias_list[idx].shape[1])
        for idx in range(len(self.m_weight_list)):
            self.m_weight_list[idx] = 1+np.random.uniform(-2,0,(self.m_weight_list[idx].shape[0],self.m_weight_list[idx].shape[1]))
        for idx in range(len(self.m_bias_list)):
            self.m_bias_list[idx] = 1+np.random.uniform(-2,0,(self.m_bias_list[idx].shape[0],self.m_bias_list[idx].shape[1]))
    def feedforward(self,a):
        for w,b in zip(self.m_weight_list, self.m_bias_list):
            a = self.m_activate_func(w.transpose() @ a + b)
        return a

    def backpropagation(self,a,output):
        nabla_w = [np.zeros(w.shape) for w in self.m_weight_list]
        nabla_b = [np.zeros(b.shape) for b in self.m_bias_list]
        # feedforward
        activation = a
        activation_list = [a]
        z_list = []
        for w, b in zip(self.m_weight_list, self.m_bias_list):
            z = w.transpose() @ activation + b
            z_list.append(z)
            activation = self.m_activate_func(z)
            activation_list.append(activation)
        # backpropagation: implement 4 equations
        # eq.1
        delta = self.m_jacobian_of_cost_func(activation_list[-1], output) * self.m_jacobian_of_activate_func(z_list[-1])
        # print('eq1 activation_list[-2].shape = ', activation_list[-2].shape)
        # print('eq1 delta.transpose().shape = ', delta.transpose().shape)
        # eq.3 and eq.4
        nabla_w[-1] = activation_list[-2] @ delta.transpose()
        nabla_b[-1] = delta
        # print('eq1 nabla_w[-1].shape = ', nabla_w[-1].shape)
        # print('eq1 nabla_b[-1].shape = ', nabla_b[-1].shape)
        # eq.2, Recursive
        for l in range(2, len(self.m_node_list)):
            a_cur = self.m_activate_func(z_list[-l])
            # print('self.m_weight_list[-l+1].shape = ', self.m_weight_list[-l+1].shape)
            # print('delta.shape = ', delta.shape)
            # print('a_cur.shape = ', a_cur.shape)
            delta = self.m_weight_list[-l+1] @ delta * a_cur
            # print('activation_list[-l-1].shape = ', activation_list[-l-1].shape)
            # print('delta.transpose().shape = ', delta.transpose().shape)
            nabla_w[-l] = activation_list[-l-1] @ delta.transpose()
            nabla_b[-l] = delta

        return (nabla_w, nabla_b)

    def evaluate(self, test_input):
        res = 0.0
        for data in test_input:
            x = data[0]
            delta = self.feedforward(x) - data[1]
            delta2 = delta ** 2
            res = res + delta2.sum()
        res = res / len(test_input)
        return res

    def updateParameters(self, mini_bach, learning_rate):
        nabla_w = [np.zeros(w.shape) for w in self.m_weight_list]
        nabla_b = [np.zeros(b.shape) for b in self.m_bias_list]
        for x, y in mini_bach:
            delta_nabla_w, delta_nabla_b = self.backpropagation(x, y)
            nabla_w = [nw + dnw for nw, dnw in zip(nabla_w, delta_nabla_w)]
            nabla_b = [nb + dnb for nb, dnb in zip(nabla_b, delta_nabla_b)]
        self.m_weight_list = [w - nw / len(mini_bach) * learning_rate for w, nw in zip(self.m_weight_list, nabla_w)]
        self.m_bias_list = [b - nb / len(mini_bach) * learning_rate  for b, nb in zip(self.m_bias_list, nabla_b)]

    def SGD(self, training_data, epochs, mini_bach_size, learning_rate, test_data=None):
        print_count = int(0.05*epochs)
        for idx in range(epochs):
            random.shuffle(training_data)
            mini_baches = [training_data[k:k+mini_bach_size] for k in range(0,len(training_data),mini_bach_size)]
            for mini_bach in mini_baches:
                self.updateParameters(mini_bach,learning_rate)
            if idx % print_count == 0:
                if test_data:
                    print("epoch {0}: {1}, size:{2}".format(idx, self.evaluate(test_data), len(test_data)))
                else:
                    print("epoch ", idx, " complete...")

def sigmoidFunc(z):
    return 1.0 / (1.0 + np.exp(-z))

def jacobianOfSigmoidFunc(z):
    return sigmoidFunc(z) * (1.0 - sigmoidFunc(z))

def ReLU(z):
    return (np.abs(z) + z) * 0.5

def jacobianOfReLU(z):
    return np.where(z > 0, 1, 0)

def jacobianOfQuadraticCostFunc(y_nn, y):
    return y_nn - y


if __name__ == '__main__':
    # node_list = [4,16,32,16,4]
    # node_list = [16,8,4,2]
    # node_list = [16,8,4]
    # node_list = [16,32,64,128,64,32,16]
    node_list = [16,32,64,32,16]
    # node_list = [16,32,16]
    train_count = 5000
    # nn = NeuralNetwork(node_list, sigmoidFunc, jacobianOfSigmoidFunc, None, jacobianOfQuadraticCostFunc)
    # nn = NeuralNetwork(node_list, ReLU, jacobianOfReLU, None, jacobianOfQuadraticCostFunc)
    nn = NeuralNetwork(node_list, sigmoidFunc, jacobianOfSigmoidFunc, None, jacobianOfQuadraticCostFunc)
    count = 100
    half_count = int(count*0.5)
    all_data_list = []
    for idx in range(count):
        input = np.random.uniform(-2,0,(node_list[0],1))+1
        output = nn.feedforward(input)
        all_data_list.append((input, output))
    training_list = all_data_list[0:half_count]
    test_list = all_data_list[half_count:count]
    # gt evalute
    gt_output = nn.evaluate(training_list)
    # init evalute
    nn.resetParameters()
    before_training_output = nn.evaluate(training_list)
    # training
    nn.SGD(training_list, train_count, int(count*0.1), 0.01, test_list)
    output = nn.evaluate(training_list)
    print('gt_output: ', gt_output)
    print('before_training_output: ', before_training_output, ', size:', len(training_list))
    print('after training output: ', output, ', size:', len(training_list))