-
Notifications
You must be signed in to change notification settings - Fork 11
/
Copy pathoptimizers.py
221 lines (156 loc) · 6.75 KB
/
optimizers.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
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
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
import numpy as np
try:
import cupy as cp
is_cupy_available = True
except:
is_cupy_available = False
from numba import njit
class SGD():
def __init__(self, alpha = 0.001):
self.alpha = alpha
def update(self, gradient, weights, v, m, v_hat, m_hat, _):
if is_cupy_available:
self._update = self._update_cupy
else:
self._update = self._update_numpy
return self._update(self.alpha, gradient, weights, v, m, v_hat, m_hat)
@staticmethod
@njit
def _update_numpy(alpha, gradient, weights, v, m, v_hat, m_hat):
weights -= gradient * alpha
return weights, v, m, v_hat, m_hat
@staticmethod
def _update_cupy(alpha, gradient, weights, v, m, v_hat, m_hat):
weights -= gradient * alpha
return weights, v, m, v_hat, m_hat
class Momentum():
def __init__(self, alpha = 0.01, beta = 0.9):
self.alpha = alpha
self.beta = beta
def update(self, gradient, weights, v, m, v_hat, m_hat, _):
if is_cupy_available:
self._update = self._update_cupy
else:
self._update = self._update_numpy
return self._update(self.alpha, self.beta, gradient, weights, v, m, v_hat, m_hat)
@staticmethod
@njit
def _update_numpy(alpha, beta, gradient, weights, v, m, v_hat, m_hat):
v = beta * v + (1 - beta) * gradient
weights -= v * alpha
return weights, v, m, v_hat, m_hat
@staticmethod
def _update_cupy(alpha, beta, gradient, weights, v, m, v_hat, m_hat):
v = beta * v + (1 - beta) * gradient
weights -= v * alpha
return weights, v, m, v_hat, m_hat
class RMSProp():
def __init__(self, alpha = 0.01, beta = 0.9, epsilon = 0.000000001):
self.alpha = alpha
self.beta = beta
self.epsilon = epsilon
def update(self, gradient, weights, v, m, v_hat, m_hat, _):
if is_cupy_available:
self._update = self._update_cupy
else:
self._update = self._update_numpy
return self._update(self.alpha, self.beta, self.epsilon, gradient, weights, v, m, v_hat, m_hat)
@staticmethod
@njit
def _update_numpy(alpha, beta, epsilon, gradient, weights, v, m, v_hat, m_hat):
v = beta * v + (1 - beta) * np.power(gradient, 2)
weights -= alpha * gradient / (np.sqrt(v) + epsilon)
return weights, v, m, v_hat, m_hat
@staticmethod
def _update_cupy(alpha, beta, epsilon, gradient, weights, v, m, v_hat, m_hat):
v = beta * v + (1 - beta) * np.power(gradient, 2)
weights -= alpha * gradient / (np.sqrt(v) + epsilon)
return weights, v, m, v_hat, m_hat
class Adam():
def __init__(self, alpha = 0.001, beta = 0.9, beta2 = 0.999, epsilon = 0.000000001):
self.alpha = alpha
self.beta = beta
self.beta2 = beta2
self.epsilon = epsilon
def update(self, gradient, weights, v, m, v_hat, m_hat, t):
if is_cupy_available:
self._update = self._update_cupy
else:
self._update = self._update_numpy
return self._update(self.alpha, self.beta, self.beta2, self.epsilon, gradient, weights, v, m, v_hat, m_hat, t)
@staticmethod
@njit
def _update_numpy(alpha, beta, beta2, epsilon, gradient, weights, v, m, v_hat, m_hat, t):
m = beta * m + (1 - beta) * gradient
v = beta2 * v + (1 - beta2) * np.power(gradient, 2)
m_hat = m / (1 - np.power(beta, t))
v_hat = v / (1 - np.power(beta2, t))
weights -= alpha * m_hat / (np.sqrt(v_hat) + epsilon)
return weights, v, m, v_hat, m_hat
@staticmethod
def _update_cupy(alpha, beta, beta2, epsilon, gradient, weights, v, m, v_hat, m_hat, t):
m = beta * m + (1 - beta) * gradient
v = beta2 * v + (1 - beta2) * np.power(gradient, 2)
m_hat = m / (1 - np.power(beta, t))
v_hat = v / (1 - np.power(beta2, t))
weights -= alpha * m_hat / (np.sqrt(v_hat) + epsilon)
return weights, v, m, v_hat, m_hat
class Nadam():
def __init__(self, alpha = 0.001, beta = 0.9, beta2 = 0.999, epsilon = 0.000000001):
self.alpha = alpha
self.beta = beta
self.beta2 = beta2
self.epsilon = epsilon
def update(self, gradient, weights, v, m, v_hat, m_hat, t):
if is_cupy_available:
self._update = self._update_cupy
else:
self._update = self._update_numpy
return self._update(self.alpha, self.beta, self.beta2, self.epsilon, gradient, weights, v, m, v_hat, m_hat, t)
@staticmethod
@njit
def _update_numpy(alpha, beta, beta2, epsilon, gradient, weights, v, m, v_hat, m_hat, t):
m = beta * m + (1 - beta) * gradient
v = beta2 * v + (1 - beta2) * np.power(gradient, 2)
m_hat = m / (1 - np.power(beta, t)) + (1 - beta) * gradient / (
1 - np.power(beta, t)
)
v_hat = v / (1 - np.power(beta2, t))
weights -= alpha * m_hat / (np.sqrt(v_hat) + epsilon)
return weights, v, m, v_hat, m_hat
@staticmethod
def _update_cupy(alpha, beta, beta2, epsilon, gradient, weights, v, m, v_hat, m_hat, t):
m = beta * m + (1 - beta) * gradient
v = beta2 * v + (1 - beta2) * np.power(gradient, 2)
m_hat = m / (1 - np.power(beta, t)) + (1 - beta) * gradient / (
1 - np.power(beta, t)
)
v_hat = v / (1 - np.power(beta2, t))
weights -= alpha * m_hat / (np.sqrt(v_hat) + epsilon)
return weights, v, m, v_hat, m_hat
class Noam():
"""Learning rate scheduler for optimizers"""
def __init__(self, optimizer, model_dim, scale_factor = 1, warmup_steps = 4000) -> None:
self.optimizer = optimizer
self.model_dim = model_dim
self.scale_factor = scale_factor
self.warmup_steps = warmup_steps
self.steps_num = 0
@staticmethod
@njit
def compute_learning_rate(scale_factor, model_dim, steps_num, warmup_steps):
return scale_factor * (
model_dim ** (-0.5)
* min(steps_num ** (-0.5), steps_num * warmup_steps ** (-1.5))
)
def update(self, gradient, weights, v, m, v_hat, m_hat, t):
self.steps_num += 1
self.optimizer.alpha = self.compute_learning_rate(self.scale_factor, self.model_dim, self.steps_num, self.warmup_steps)
return self.optimizer.update(gradient, weights, v, m, v_hat, m_hat, t)
optimizers = {
"sgd": SGD(),
"momentum": Momentum(),
"rmsprop": RMSProp(),
"adam": Adam(),
"nadam": Nadam(),
}