Neural-Network-From-Scratch/neural_net.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
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
198
199
200
201
202
203
import numpy as np
from layer import Layer, InputLayer
from sklearn.datasets import fetch_openml, load_iris
from sklearn.preprocessing import OneHotEncoder, LabelEncoder
from sklearn.model_selection import train_test_split
from activations import *
from tqdm import tqdm

class NeuralNetwork():
    def __init__(self, optimizer="adam", eta=.01, alpha=.9, beta1=.9, beta2=.999):
        self.layers = []
        self.optimizer = optimizer
        self.eta = eta
        self.alpha = alpha
        self.beta1 = beta1
        self.beta2 = beta2


    def add_layer(self, layer):
        if len(self.layers) > 0:
            input_dim = self.layers[len(self.layers)-1].output_dim
            layer.update_dim(input_dim)

        self.layers.append(layer)


    def error(self, expected, predicted):
        error = np.subtract(expected, predicted)
        squared = .5 * np.square(error)
        error = np.sum(squared)
        return error


    def feedforward(self, inputs):
        self.layers[0].activations = inputs
        for i in range(1, len(self.layers)):
            layer_i = self.layers[i]
            layer_i.feedforward(self.layers[i-1])

        output = self.layers[len(self.layers)-1].activations

        return output # return the output after the inputs feedforward through the network


    def calc_layer_errors(self, e_out):
        for i in range(len(self.layers)-1, 0, -1):
            layer_i = self.layers[i]
            if isinstance(layer_i, ConvalutionLayer):
                prev_layer = self.layers[i+1]
                if isinstance(prev_layer, ConvalutionLayer):
                    prev_kernel_size = prev_layer.kernel_size
                    prev_kernel = prev_layer.kernel
                    prev_cache = prev_layer.cache
                    prev_error = prev_layer.error
                    prev_activations = layer_i.activations
                    for m in range(layer_i.output_dim):
                        for n in range(layer_i.output_dim):
                            for a in range(prev_kernel_size):
                                for b in range (prev_kernel_size):
                                    layer_i.error[m, n] = prev_error[m-a, n-b]*sigmoid_der(prev_activations[m-a, n-b])*prev_kernel[a, b]

                elif isinstance(prev_layer, Layer):
                    prev_error = prev_layer.error
                    prev_weights = prev_layer.weights
                    sum_weights = prev_weights.sum(axis=1)
                    a = prev_weights.transpose()
                    b = np.dot(a, prev_errors)
                    layer_i.error = b.reshape(layer_i.output_dim, layer_i.output_dim)

            elif isinstance(layer_i, Layer):
                if i == len(self.layers)-1:
                    layer_i.error = e_out
                else:
                    prev_layer = self.layers[i+1]
                    prev_weights = prev_layer.weights
                    prev_errors = prev_layer.error
                    sum_weights = prev_weights.sum(axis=1)
                    a = prev_weights.transpose()

                    b = np.dot(a, prev_errors)

                    layer_i.error = b


    def backprop(self):
        for i in range(len(self.layers)-1, 0, -1):
            layer_i = self.layers[i]
            predicted_i = layer_i.activations
            cache_i = layer_i.cache
            error_i = layer_i.error
            activation_i = layer_i.activation

            if activation_i is "sig":
                delta_weights = np.dot((-error_i * sigmoid_der(predicted_i)).reshape((-1, 1)), self.layers[i-1].activations.reshape((1, -1)))
                delta_biases = -error_i * sigmoid_der(predicted_i)
            elif activation_i is "relu":
                delta_weights = np.dot((-error_i * relu_der(cache_i)).reshape((-1, 1)), self.layers[i-1].activations.reshape((1, -1)))
                delta_biases = -error_i * relu_der(cache_i)
            elif activation_i is "leaky":
                delta_weights = np.dot((-error_i * leaky_relu_der(cache_i)).reshape((-1, 1)), self.layers[i-1].activations.reshape((1, -1)))
                delta_biases = -error_i * leaky_relu_der(cache_i)
            elif activation_i is "linear":
                delta_weights = np.dot((-error_i * linear(cache_i)).reshape((-1, 1)), self.layers[i-1].activations.reshape((1, -1)))
                delta_biases = -error_i * linear_der(cache_i)

            layer_i.add_gradient(delta_weights)

            if self.optimizer is "adam":
                layer_i.adam(delta_weights, delta_biases, self.beta1, self.beta2, self.eta)
            elif self.optimizer is "rmsprop":
                layer_i.rmsprop(delta_weights, delta_biases, self.eta, self.alpha)
            elif self.optimizer is "grad":
                layer_i.grad(delta_weights, delta_biases, self.eta)


    def fit(self, data, labels, epochs=1):
        print("Training...")
        for epoch in range(epochs):
            print("Epoch " + str(epoch + 1))
            for i in tqdm(range(len(data))):
                data_i = data[i]
                label_i = labels[i]
                out = self.feedforward(data_i)
                self.calc_layer_errors(label_i - out)
                self.backprop()
        print("Finished training!")


    def predict(self, x):
        return self.feedforward(x)


    def evaluate(self, X, y):
        correct = 0
        print("Evaluating...")
        for i in range(0, len(X)):
            predicted = self.predict(X[i])
            max_value = np.amax(predicted)
            predicted = np.array([0 if a < max_value else 1 for a in predicted])

            if np.array_equal(predicted, y[i]):
                correct += 1
        print("Finished evaluating!")
        return correct/len(X)


# ohe = OneHotEncoder(sparse=False)

# mnist = fetch_openml('mnist_784', version=1)
# X, y = mnist["data"], mnist["target"]

# y = ohe.fit_transform(y.reshape((-1, 1)))

# X_train, X_test, y_train, y_test = X[:60000], X[60000:], y[:60000], y[60000:]
# X_train = np.array(X_train)
# X_test = np.array(X_test)
# y_train = np.array(y_train)
# y_test = np.array(y_test)

# iris = load_iris()

# X = iris.data
# y = iris.target

# y = ohe.fit_transform(y.reshape((-1, 1)))

# X_train, X_test, y_train, y_test=train_test_split(X,y,test_size=.5)

# def f(x):
#     return x**2 + 2*x + 3

# X = []
# y = []
# for i in range(-1000, 1000, 2):
#     a = np.random.rand()/10
#     X.append(np.array([i*a]).reshape(1, ))
#     y.append(np.array([f(i*a)]).reshape(1, ))

# X = np.array(X)
# y = np.array(y)

# N = NeuralNetwork(eta=.001, optimizer="adam")
# N.add_layer(InputLayer(784))
# N.add_layer(Layer(128, activation='leaky'))
# N.add_layer(Layer(32, activation='leaky'))
# N.add_layer(Layer(16, activation='leaky'))
# N.add_layer(Layer(10, activation='sig'))

# N.fit(X_train, y_train, epochs=10)

# # for i in range(-100, 100, 3):
# #     print(str(N.predict(np.array([i]).reshape(1, ))) + " " + str(f(i)))

# print(N.evaluate(X_test, y_test))

network = NeuralNetwork()
network.add_layer(InputLayer(3))
network.add_layer(Layer(2))
network.add_layer(Layer(1))

result = network.feedforward(np.array([1, 1, 1]))
print(result)