Neural-Network-From-Scratch/layer.py at master · ethan-stone/Neural-Network-From-Scratch · 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
from activations import *
import numpy as np

class Layer():
    def __init__(self, output_dim, activation="sig"):
        """
        Constructor

        input_dim : number of inputs for the layer (optional)
        output_dim : number of outputs for the layer, essentially the number of neurons in the layer

        """
        self.input_dim = 0
        self.output_dim = output_dim
        self.activations = np.array([])
        self.activation = activation
        self.weights = np.random.randn(self.output_dim, self.input_dim)
        self.biases = np.zeros(output_dim)
        self.cache = 0
        self.error = np.zeros(output_dim)
        self.gradients = np.zeros((self.output_dim, self.input_dim))
        self.epsilon = np.full((self.output_dim, self.input_dim), 1)*(10**(-8))
        self.prev_E = 0
        self.prev_m = 0
        self.prev_v = 0
        self.t = 1


    def __str__(self):
        return f"Layer -> input_dim : {self.input_dim}, output_dim : {self.output_dim}"


    def update_dim(self, input_dim):
        self.input_dim = input_dim
        if self.activation is "sig":
            self.weights = np.random.randn(self.output_dim, self.input_dim)*(4*np.sqrt(2/(self.input_dim + self.output_dim)))
        elif self.activation is "relu" or self.activation is "leaky" or self.activation is "linear":
            self.weights = np.random.randn(self.output_dim, self.input_dim)*(np.sqrt(2)*np.sqrt(2/(self.input_dim + self.output_dim)))
        self.gradients = np.zeros((self.output_dim, self.input_dim))
        self.epsilon = np.full((self.output_dim, self.input_dim), 1)*(10**(-8))


    def add_gradient(self, gradient):
        self.gradients = self.gradients + gradient**2


    def E(self, x, alpha):
        E = (1-alpha)*x**2 + alpha*self.prev_E
        self.prev_E = E
        return E


    def rmsprop(self, delta_weights, delta_biases, eta, alpha):
        self.weights = self.weights - (eta * delta_weights)/np.sqrt(self.epsilon + self.E(delta_weights, alpha))
        self.biases = self.biases - (eta * delta_biases)


    def adam(self, delta_weights, delta_biases, beta1, beta2, eta):
        m = beta1 * self.prev_m + (1 - beta1) * delta_weights
        v = beta2 * self.prev_v + (1 - beta2) * (delta_weights**2)

        self.prev_m = m
        self.prev_v = v

        m_hat = m/(1 - beta1**self.t)
        v_hat = v/(1 - beta2**self.t)

        self.t += 1
        self.weights = self.weights - (eta * m_hat)/(self.epsilon + np.sqrt(v_hat))
        self.biases = self.biases - (eta * delta_biases)


    def grad(self, delta_weights, delta_biases, eta):
        self.weights = self.weights - (eta * delta_weights)
        self.biases = self.biases - (eta * delta_biases)


    def feedforward(self, input_layer):
        input_activations = input_layer.activations
        if isinstance(input_layer, Layer):
            input_activations = input_activations.flatten()

        dot_product = np.dot(self.weights, input_activations)
        activations = np.add(dot_product, self.biases)
        self.cache = activations

        if self.activation is 'sig':
            self.activations = sigmoid(activations)
        elif self.activation is 'relu':
            self.activations = relu(activations)
        elif self.activation is 'leaky':
            self.activations = leaky_relu(activations)
        elif self.activation is 'linear':
            self.activations = linear(activations)


class InputLayer(Layer):
    def __init__(self, input_dim):
        self.input_dim = input_dim
        self.output_dim = input_dim
        self.activations = np.array([])


    def __str__(self):
        return f"InputLayer -> input_dim : {self.input_dim}"


class ConvulutionalLayer():
    def __init__(self, kernel_size=3, stride=1, padding=0, activation="relu"):
        self.activation = activation
        self.input_dim = 0
        self.kernel_size = kernel_size
        self.stride = stride
        self.padding = padding
        self.output_dim = 0
        self.kernel = np.random.randn(self.kernel_size, self.kernel_size)
        self.cache = np.zeros((self.output_dim, self.output_dim))
        self.activations = np.zeros((self.output_dim, self.output_dim))
        self.error = 0


    def update_dim(self, input_dim):
        self.input_dim = input_dim
        self.output_dim = (self.input_dim - self.kernel_size + 2*self.padding)/self.stride + 1
        self.cache = np.zeros((self.output_dim, self.output_dim))
        self.activations = np.zeros((self.output_dim, self.output_dim))


    def feedforward(self, input_layer):
        input_activations = input_layer.activations

        for i in range(self.output_dim):
            for j in range(self.output_dim):
                chunk = input_activations[i:i+self.kernel_size, j:j+self.kernel_size]
                feature_cache = np.multiply(self.kernel, chunk).sum()

                if self.activation is 'sig':
                    feature = sigmoid(feature_cache)
                elif self.activation is 'relu':
                    feature = relu(feature_cache)
                elif self.activation is 'leaky':
                    feature = leaky(feature_cache)
                elif self.activation is 'linear':
                    feature = linear(feature_cache)

                self.cache[i, j] = feature_cache
                self.activations[i,j] = feature