Neural-Network/NeuralNetwork.py at main · YonatanVigiser/Neural-Network · 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
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
import random
import numpy as np
from tqdm import tqdm


class NeuralNetwork:
    def __init__(self, neurons_count: list[int]):
        self.neurons_count = neurons_count
        self.hidden_layers_count = len(neurons_count) - 2

        self.weights = self.generate_random_weights()
        self.biases = self.generate_random_biases()

    @staticmethod
    def sig(x: float) -> float:
        return 1 / (1 + np.exp(-x))

    @staticmethod
    def cost(desired_values: list[float], values: list[float]) -> float:
        res = 0
        for i in range(len(desired_values)):
            res += (values[i] - desired_values[i]) ** 2
        return res

    # Impacts:
    @staticmethod
    def impact_of_output_weight(activation: float, desired_activation: float, activation_before: float) -> float:
        return 2 * (activation - desired_activation) * activation * (1 - activation) * activation_before

    @staticmethod
    def impact_of_output_bias(activation: float, desired_value: float) -> float:
        return 2 * (activation - desired_value) * activation * (1 - activation)

    @staticmethod
    def impact_of_output_activation(activation: float, desired_values: list[float], weights: list[float]) -> float:
        return sum([2 * (activation - desired_values[i]) * activation * (1 - activation) *
                    weights[i] for i in range(len(weights))])

    @staticmethod
    def impact_of_weight(activation: float, activation_before: float, layer_after_impact: float) -> float:
        return activation * (1 - activation) * activation_before * layer_after_impact

    @staticmethod
    def impact_of_bias(activation: float, this_layer_impact: float) -> float:
        return activation * (1 - activation) * this_layer_impact

    @staticmethod
    def impact_of_activation(activation: float, weights: list[float], next_layer_impact: list[float]) -> float:
        return sum([activation * (1 - activation) * weights[i] * next_layer_impact[i] for i in range(len(weights))])

    def weights_impact_one_layer(self, network_activations: list[list[float]], layer: int, desired_output: list[float],
                                 next_layer_impact: list[float]) -> list[list[float]]:
        layer_activation = network_activations[layer]
        layer_before_activation = network_activations[layer - 1]

        res = [[] for _ in range(len(layer_activation))]

        if len(next_layer_impact) == 0:
            for i in range(len(layer_activation)):
                for x in range(len(layer_before_activation)):
                    res[i].append(round(self.impact_of_output_weight(layer_activation[i], desired_output[i],
                                                                     layer_before_activation[x]), 15))
        else:
            for i in range(len(layer_activation)):
                for x in range(len(layer_before_activation)):
                    res[i].append(
                        round(self.impact_of_weight(layer_activation[i], layer_before_activation[x],
                                                    next_layer_impact[i]), 15))

        return res

    def bias_impact_one_layer(self, network_activations: list[list[float]], layer: int, desired_output: list[float],
                              this_layer_impact: list[float]) -> list[float]:
        layer_activation = network_activations[layer]
        res = []

        if len(this_layer_impact) == 0:
            for i in range(len(layer_activation)):
                res.append(self.impact_of_output_bias(layer_activation[i], desired_output[i]))

        else:
            for i in range(len(layer_activation)):
                res.append(self.impact_of_bias(layer_activation[i], this_layer_impact[i]))

        return res

    def activation_impact_one_layer(self, network_activations: list[list[float]], layer: int,
                                    desired_outputs: list[float], normal_weights, next_layer_impact):
        layer_activation = network_activations[layer - 1]
        res = []
        weights = [[] for _ in range(len(layer_activation))]

        for i in range(len(normal_weights[layer - 1])):
            for x in range(len(normal_weights[layer - 1][i])):
                weights[x].append(normal_weights[layer - 1][i][x])

        if len(next_layer_impact) == 0:
            for i in range(len(layer_activation)):
                res.append(round(self.impact_of_output_activation(
                    layer_activation[i], desired_outputs, weights[i]), 15))
        else:
            for i in range(len(layer_activation)):
                res.append(round(self.impact_of_activation(layer_activation[i], weights[i], next_layer_impact), 15))

        return res

    def generate_random_weights(self) -> list[list[list[float]]]:
        weights = [[] for _ in range(self.hidden_layers_count + 1)]

        for i in range(1, len(self.neurons_count)):
            for x in range(self.neurons_count[i]):
                weights[i - 1].append([round(random.uniform(-1, 1), 10) for _ in range(self.neurons_count[i - 1])])

        return weights

    def generate_random_biases(self) -> list[list[float]]:
        biases = [[] for _ in range(self.hidden_layers_count + 1)]

        for i in range(1, len(self.neurons_count)):
            for _ in range(self.neurons_count[i]):
                biases[i - 1].append(round(random.uniform(-0.5, 0.5), 10))

        return biases

    def generate_biases_zero(self) -> list[list[float]]:
        biases = [[] for _ in range(self.hidden_layers_count + 1)]

        for i in range(1, len(self.neurons_count)):
            for _ in range(self.neurons_count[i]):
                biases[i - 1].append(0)
        return biases

    def run_one_layer(self, layer_num: int, inputs: list[float]) -> list[float]:
        output = []
        weights = self.weights[layer_num]
        biases = self.biases[layer_num]

        for i in range(len(biases)):
            res = biases[i]
            for x in range(len(weights[i])):
                res += weights[i][x] * inputs[x]
            res = float(self.sig(res))
            output.append(res)

        return output

    def run_network(self, inputs: list[float]) -> list[list[float]]:
        result = [inputs]  # added inputs directly to history
        for i in range(len(self.weights)):
            layer = self.run_one_layer(i, inputs)
            inputs = layer

            result.append(layer)

        return result

    def train(self, training_inputs: list[list[float]], training_outputs: list[int], learning_rate: float) -> None:
        print("Training started!")
        for i in tqdm(range(len(training_outputs))):
            inputs = training_inputs[i]
            desired_output = training_outputs[i]
            network = self.run_network(inputs)
            # network.insert(0, inputs)

            impact = [[] for i in range(self.hidden_layers_count + 3)]
            desired_values = [0] * self.neurons_count[-1]
            desired_values[desired_output] = 1

            for layer in range(self.hidden_layers_count + 1, 0, -1):
                weights_impact = self.weights_impact_one_layer(network, layer, desired_values, impact[layer + 1])
                impact[layer] = self.activation_impact_one_layer(network, layer, desired_values,
                                                                 self.weights, impact[layer + 1])

                if layer == self.hidden_layers_count + 1:
                    biases_impact = self.bias_impact_one_layer(network, layer, desired_values, [])
                else:
                    biases_impact = self.bias_impact_one_layer(network, layer, desired_values, impact[layer])

                for n in range(len(weights_impact)):
                    self.biases[layer - 1][n] = self.biases[layer - 1][n] - biases_impact[n] * learning_rate
                    for x in range(len(weights_impact[n])):
                        self.weights[layer - 1][n][x] = (self.weights[layer - 1][n][x] - weights_impact[n][x] *
                                                         learning_rate)
        print("\nFinished Training!\n")

    def test(self, training_inputs: list[list[float]], training_outputs: list[int]) -> float:
        print("Test started!")
        correct = 0
        for i in tqdm(range(len(training_outputs))):
            network = self.run_network(training_inputs[i])
            prediction = network[-1].index(max(network[-1]))
            if prediction == training_outputs[i]:
                correct += 1

        percentage = (100 * correct) / len(training_outputs)
        print(f"\nFinished testing with {percentage}% success rate!")
        return percentage