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
| import numpy as np
class ANN:
def __init__(self, layers):
self.W = []
self.b = []
for x_nodes, h_nodes in layers:
self.add_layer(x_nodes, h_nodes)
def add_layer(self, x_nodes, h_nodes):
self.W.append(np.random.rand(x_nodes, h_nodes))
self.b.append(np.random.rand(h_nodes))
def feed_forward(self):
delta = 1e-7
a = self.x
for i, w in enumerate(self.W):
b = self.b[i]
z = np.dot(a, w) + b
a = self.activate(z)
y = a
return -np.sum(self.t*np.log(y + delta) + (1-self.t) * np.log((1 - y) + delta))
def train(self, x, t, epochs=1, learning_rate=1e-2, activate=None, debug_step=None):
self.x = x
self.t = t
self.activate = activate or ANN.sigmoid
if not debug_step:
debug_step = int(epochs * 0.1)
f = lambda x: self.feed_forward()
for step in range(epochs):
for i in range(len(self.W)):
self.W[i] -= learning_rate * ANN.derivative(f, self.W[i])
self.b[i] -= learning_rate * ANN.derivative(f, self.b[i])
if step % debug_step == 0:
print('step = ', step, 'loss value = ', self.feed_forward())
def predict(self, input_data):
a = input_data
for i, w in enumerate(self.W):
b = self.b[i]
z = np.dot(a, w) + b
a = self.activate(z)
return a
@staticmethod
def derivative(f, x, dx=1e-5):
grad = np.zeros_like(x)
it = np.nditer(x, flags=['multi_index'], op_flags=['readwrite'])
while not it.finished:
idx = it.multi_index
tmp_val = x[idx]
x[idx] = tmp_val + dx
fx1 = f(x)
x[idx] = tmp_val - dx
fx2 = f(x)
grad[idx] = (fx1 - fx2) / (2 * dx)
x[idx] = tmp_val
it.iternext()
return grad
@staticmethod
def sigmoid(z):
return 1 / (1+np.exp(z))
|
