-
Notifications
You must be signed in to change notification settings - Fork 0
/
optimizer.py
180 lines (138 loc) · 6.6 KB
/
optimizer.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
"""
TODO:
- add weight decay
"""
import torch
class SGD(torch.optim.Optimizer):
def __init__(self, params, lr=0.001):
defaults={'lr':lr}
super().__init__(params, defaults)
def step(self, closure=None):
loss = None
if closure is not None:
loss = closure()
for group in self.param_groups:
for p in group['params']:
p.data.add_(-group['lr']*p.grad.data)
return loss
class Momentum(torch.optim.Optimizer):
def __init__(self, params, lr=0.001, momentum=0.8, dampening=0.2):
self.momentum = momentum
self.dampening = dampening
defaults = {'lr': lr}
super().__init__(params, defaults)
def step(self):
for group in self.param_groups:
for idx, p in enumerate(group['params']):
if idx not in self.state['running_grad']:
self.state['running_grad'][idx] = p.grad.data
else:
self.state['running_grad'][idx].add_(
self.state['running_grad'][idx].data * self.momentum + \
p.grad.data * (1 - self.dampening)
)
p.data.add_(-group['lr'] * self.state['running_grad'][idx])
class Nestrov(torch.optim.Optimizer):
"""
torch naive implementation:
----
b_t = momentum * b_t-1 + g_t
b_t = momentum * b_t + g_t
w = w - lr * b_t
another implementation:
----
b_t = momentum * b_t-1 + dampening * g_t
w = w - momentum * b_t - lr * g_t
"""
def __init__(self, params, lr, momentum=0.8):
self.momentum = momentum
defaults = {'lr': lr}
super().__init__(params, defaults)
def step(self):
for group in self.param_groups:
for idx, p in enumerate(group['params']):
if idx not in self.state['running_grad']:
self.state['running_grad'][idx] = p.grad.data
else:
self.state['running_grad'][idx].add_(
self.state['running_grad'][idx].data * self.momentum + \
p.grad.data
)
grad = self.momentum * self.state['running_grad'][idx] + p.grad.data
p.data.add_(-group['lr'] * grad)
class Adam(torch.optim.Optimizer):
def __init__(self, params, lr=0.001, beta1=0.9, beta2=0.999, eps=1e-8):
defaults = {'lr': lr}
self.beta1 = beta1
self.beta2 = beta2
self.eps = eps
super().__init__(params, defaults)
def step(self):
for group in self.param_groups:
for idx, p in enumerate(group['params']):
# first moment
if idx not in self.state['runing_1-th_moment']:
self.state['runing_1-th_moment'][idx] = p.grad.data
else:
self.state['runing_1-th_moment'][idx] = self.beta1 * \
self.state['runing_1-th_moment'][idx] + (1 - self.beta1) * p.grad.data
# second moment
if idx not in self.state['runing_2-th_moment']:
self.state['runing_2-th_moment'][idx] = p.grad.data ** 2
else:
self.state['runing_2-th_moment'][idx] = self.beta1 * \
self.state['runing_2-th_moment'][idx] + (1 - self.beta1) * p.grad.data ** 2
first_moment = self.state['runing_1-th_moment'][idx] / (1 - self.beta1)
second_moment = self.state['runing_2-th_moment'][idx] / (1-self.beta2)
p.data.add_(-group['lr'] * (first_moment / (torch.sqrt(second_moment) + self.eps)))
class Nadam(torch.optim.Optimizer):
def __init__(self, params, lr=0.001, beta1=0.9, beta2=0.999, eps=1e-8):
defaults = {'lr': lr}
self.beta1 = beta1
self.beta2 = beta2
self.eps = eps
super().__init__(params, defaults)
def step(self):
for group in self.param_groups:
for idx, p in enumerate(group['params']):
# first moment
if idx not in self.state['runing_1-th_moment']:
self.state['runing_1-th_moment'][idx] = p.grad.data
else:
self.state['runing_1-th_moment'][idx] = self.beta1 * \
self.state['runing_1-th_moment'][idx] + (1 - self.beta1) * p.grad.data
# second moment
if idx not in self.state['runing_2-th_moment']:
self.state['runing_2-th_moment'][idx] = p.grad.data ** 2
else:
self.state['runing_2-th_moment'][idx] = self.beta1 * \
self.state['runing_2-th_moment'][idx] + (1 - self.beta1) * p.grad.data ** 2
first_moment = self.beta1 * self.state['runing_1-th_moment'][idx] / (1 - self.beta1) + p.grad.data
second_moment = self.state['runing_2-th_moment'][idx] / (1-self.beta2)
p.data.add_(-group['lr'] * (first_moment / (torch.sqrt(second_moment) + self.eps)))
class Adamw(torch.optim.Optimizer):
def __init__(self, params, lr=0.001, beta1=0.9, beta2=0.999, eps=1e-8):
defaults = {'lr': lr}
self.beta1 = beta1
self.beta2 = beta2
self.eps = eps
super().__init__(params, defaults)
def step(self):
for group in self.param_groups:
for idx, p in enumerate(group['params']):
p.data.add_(-group['lr'] * p.data)
# first moment
if idx not in self.state['runing_1-th_moment']:
self.state['runing_1-th_moment'][idx] = p.grad.data
else:
self.state['runing_1-th_moment'][idx] = self.beta1 * \
self.state['runing_1-th_moment'][idx] + (1 - self.beta1) * p.grad.data
# second moment
if idx not in self.state['runing_2-th_moment']:
self.state['runing_2-th_moment'][idx] = p.grad.data ** 2
else:
self.state['runing_2-th_moment'][idx] = self.beta1 * \
self.state['runing_2-th_moment'][idx] + (1 - self.beta1) * p.grad.data ** 2
first_moment = self.state['runing_1-th_moment'][idx] / (1 - self.beta1)
second_moment = self.state['runing_2-th_moment'][idx] / (1 - self.beta2)
p.data.add_(-group['lr'] * (first_moment / (torch.sqrt(second_moment) + self.eps)))