-
Notifications
You must be signed in to change notification settings - Fork 11
/
Copy pathlayer_norm.py
133 lines (88 loc) · 4.54 KB
/
layer_norm.py
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
try:
import cupy as np
is_cupy_available = True
except:
import numpy as np
is_cupy_available = False
class LayerNormalization():
"""
Applies layer normalization to the input data
---------------------------------------------
Args:
`momentum` (float): the momentum parameter of the moving mean
`epsilon` (float): the epsilon parameter of the algorithm
Returns:
output: the normalized input data with same shape
"""
def __init__(self, normalized_shape = None, epsilon = 0.001, data_type = np.float32):
self.normalized_shape = normalized_shape
self.normalized_axis = None
self.epsilon = epsilon
self.gamma = None
self.beta = None
self.mean = None
self.var = None
self.optimizer = None
self.data_type = data_type
self.axis = None
self.build()
def set_optimizer(self, optimizer):
self.optimizer = optimizer
def build(self):
self.feature_size = None
if self.normalized_shape is not None:
self.gamma = np.ones(self.normalized_shape).astype(self.data_type)
self.beta = np.zeros(self.normalized_shape).astype(self.data_type)
self.vg, self.mg = np.zeros_like(self.gamma).astype(self.data_type), np.zeros_like(self.gamma).astype(self.data_type)
self.vg_hat, self.mg_hat = np.zeros_like(self.gamma).astype(self.data_type), np.zeros_like(self.gamma).astype(self.data_type)
self.vb, self.mb = np.zeros_like(self.gamma).astype(self.data_type), np.zeros_like(self.gamma).astype(self.data_type)
self.vb_hat, self.mb_hat = np.zeros_like(self.gamma).astype(self.data_type), np.zeros_like(self.gamma).astype(self.data_type)
def forward(self, X):
self.input_data = X
x_T = self.input_data.T
if self.normalized_shape is None:
self.normalized_shape = self.input_data.shape[1:]
self.build()
self.normalized_axis = tuple(np.arange(self.input_data.ndim - self.gamma.ndim).tolist())
self.feature_size = self.gamma.size
self.mean = np.mean(x_T, axis = 0)
self.var = np.var(x_T,axis = 0)
self.X_centered = (x_T - self.mean)
self.stddev_inv = 1 / np.sqrt(self.var + self.epsilon)
self.X_hat_T = self.X_centered * self.stddev_inv
self.X_hat = self.X_hat_T.T
self.output_data = self.gamma * self.X_hat + self.beta
return self.output_data
def backward(self, error):
error_T = error.T
#first variant
output_error = (1 / self.feature_size) * np.expand_dims(self.gamma, axis = self.normalized_axis).T * self.stddev_inv * (#self.gamma[np.newaxis, :].T
self.feature_size * error_T
- np.sum(error_T, axis = 0)
- self.X_centered * np.power(self.stddev_inv, 2) * np.sum(error_T * self.X_centered, axis = 0)
)
#second variant
# dX_hat = error_T * self.gamma[np.newaxis, :].T
# output_error = (1 / self.feature_size) * self.stddev_inv * (
# self.feature_size * dX_hat
# - np.sum(dX_hat, axis = 0)
# - self.X_hat_T * np.sum(dX_hat * self.X_hat_T, axis = 0)
# )
#third (naive slow) variant
# x_T = self.input_data.T
# dX_hat = error_T * self.gamma[np.newaxis, :].T
# dvar = np.sum(dX_hat * self.X_centered, axis=0) * -.5 * self.stddev_inv**3
# dmu = np.sum(dX_hat * -self.stddev_inv, axis=0) + dvar * np.mean(-2. * self.X_centered, axis=0)
# output_error = (dX_hat * self.stddev_inv) + (dvar * 2 * self.X_centered / self.feature_size) + (dmu / self.feature_size)
output_error = output_error.T
self.grad_gamma = np.sum(error * self.X_hat, axis = self.normalized_axis)
self.grad_beta = np.sum(error, axis = self.normalized_axis)
return output_error
def update_weights(self, layer_num):
self.gamma, self.vg, self.mg, self.vg_hat, self.mg_hat = self.optimizer.update(self.grad_gamma, self.gamma, self.vg, self.mg, self.vg_hat, self.mg_hat, layer_num)
self.beta, self.vb, self.mb, self.vb_hat, self.mb_hat = self.optimizer.update(self.grad_beta, self.beta, self.vb, self.mb, self.vb_hat, self.mb_hat, layer_num)
return layer_num + 1
def get_grads(self):
return self.grad_gamma, self.grad_beta
def set_grads(self, grads):
self.grad_gamma, self.grad_beta = grads