deep_learning_basics/neural_net.py at master · lukDev/deep_learning_basics · 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
from backprop import *
import numpy as np


class NeuralNet:
    def __init__(self, layers, loss_func, loss_func_der, out_interpretation):
        self.layers = layers
        self.loss_func = loss_func
        self.loss_func_der = loss_func_der
        self.out_interpretation = out_interpretation

    def to_graph(self, with_loss):
        nodes = [Node(None, None, (self.input_dimension(), 1))]  # input node
        edges = []

        for layer in self.layers:
            prev_out_node = nodes[-1]

            def weights_func(w):
                return lambda _: w
            weight_node = Node(weights_func(layer.weights), None, (layer.output_dimension(), layer.input_dimension()))
            nodes.append(weight_node)

            def matmul(prev_activations):  # matrix multiplication
                return np.matmul(prev_activations[0], prev_activations[1])

            def matmul_grad(forward_grad, prev_activations):  # gradient for matrix multiplication
                return (
                    np.matmul(forward_grad, np.transpose(prev_activations[1])),
                    np.matmul(np.transpose(prev_activations[0]), forward_grad)
                )

            matmul_node = Node(matmul, matmul_grad, (layer.output_dimension(), 1))
            nodes.append(matmul_node)

            def bias_func(b):
                return lambda _: b
            bias_node = Node(bias_func(layer.biases), None, (layer.output_dimension(), 1))
            nodes.append(bias_node)

            def plus(prev_activations):
                return np.add(prev_activations[0], prev_activations[1])

            def plus_grad(forward_grad, _):
                return forward_grad, forward_grad

            plus_node = Node(plus, plus_grad, (layer.output_dimension(), 1))
            nodes.append(plus_node)

            def get_activation_func(act_func):
                def activation(prev_activations):
                    prev_result = prev_activations[0]
                    act_res = act_func(prev_result)
                    return act_res

                return activation

            def get_act_der(act_der):
                def activation_der(forward_grad, prev_activations):
                    prev_result = prev_activations[0]
                    derived_acts = act_der(prev_result)
                    new_grad = np.multiply(forward_grad, derived_acts)
                    return (new_grad,)

                return activation_der

            activation_node = Node(get_activation_func(layer.activation_func), get_act_der(layer.act_func_der), (layer.output_dimension(), 1))
            nodes.append(activation_node)

            edges.append((weight_node, matmul_node))
            edges.append((prev_out_node, matmul_node))
            edges.append((matmul_node, plus_node))
            edges.append((bias_node, plus_node))
            edges.append((plus_node, activation_node))

        if with_loss:
            y_head_node = nodes[-1]
            target_node = Node(None, None, (self.output_dimension(), 1))
            nodes.append(target_node)

            def loss(prev_activations):
                return self.loss_func(prev_activations[0], prev_activations[1])

            def loss_der(_, prev_activations):
                return (
                    self.loss_func_der(prev_activations[0], prev_activations[1]),
                    0  # w.r.t. target --> not needed
                )

            loss_node = Node(loss, loss_der, (1, 1))
            nodes.append(loss_node)

            edges.append((y_head_node, loss_node))
            edges.append((target_node, loss_node))

            pois_w = [i * 5 + 1 for i in range(len(self.layers))]
            pois_b = [i * 5 + 3 for i in range(len(self.layers))]
            pois = [None] * (2 * len(self.layers))
            pois[::2] = pois_w
            pois[1::2] = pois_b

            ins = [0, len(nodes) - 2]

            return Graph(nodes, edges, ins, pois)

        else:
            return Graph(nodes, edges, [0], [])

    def input_dimension(self):
        return self.layers[0].input_dimension()

    def output_dimension(self):
        return self.layers[-1].output_dimension()

    def evaluate(self, x):
        graph = self.to_graph(False)
        return self.out_interpretation(graph.evaluate([x])[graph.nodes[-1]])


class Layer:
    def __init__(self, weights, biases, activation_func, act_func_der):
        self.weights = weights
        self.biases = biases
        self.activation_func = activation_func
        self.act_func_der = act_func_der

    def input_dimension(self):
        if len(self.weights) == 0:
            return None

        return len(self.weights[0])

    def output_dimension(self):
        return len(self.biases)