-
Notifications
You must be signed in to change notification settings - Fork 334
/
Copy pathadabound.py
234 lines (202 loc) · 11.1 KB
/
adabound.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
222
223
224
225
226
227
228
229
230
231
232
233
234
import math
import torch
from torch.optim import Optimizer
class AdaBound(Optimizer):
"""Implements AdaBound algorithm.
It has been proposed in `Adaptive Gradient Methods with Dynamic Bound of Learning Rate`_.
Arguments:
params (iterable): iterable of parameters to optimize or dicts defining
parameter groups
lr (float, optional): Adam learning rate (default: 1e-3)
betas (Tuple[float, float], optional): coefficients used for computing
running averages of gradient and its square (default: (0.9, 0.999))
final_lr (float, optional): final (SGD) learning rate (default: 0.1)
gamma (float, optional): convergence speed of the bound functions (default: 1e-3)
eps (float, optional): term added to the denominator to improve
numerical stability (default: 1e-8)
weight_decay (float, optional): weight decay (L2 penalty) (default: 0)
amsbound (boolean, optional): whether to use the AMSBound variant of this algorithm
.. Adaptive Gradient Methods with Dynamic Bound of Learning Rate:
https://openreview.net/forum?id=Bkg3g2R9FX
"""
def __init__(self, params, lr=1e-3, betas=(0.9, 0.999), final_lr=0.1, gamma=1e-3,
eps=1e-8, weight_decay=0, amsbound=False):
if not 0.0 <= lr:
raise ValueError("Invalid learning rate: {}".format(lr))
if not 0.0 <= eps:
raise ValueError("Invalid epsilon value: {}".format(eps))
if not 0.0 <= betas[0] < 1.0:
raise ValueError("Invalid beta parameter at index 0: {}".format(betas[0]))
if not 0.0 <= betas[1] < 1.0:
raise ValueError("Invalid beta parameter at index 1: {}".format(betas[1]))
if not 0.0 <= final_lr:
raise ValueError("Invalid final learning rate: {}".format(final_lr))
if not 0.0 <= gamma < 1.0:
raise ValueError("Invalid gamma parameter: {}".format(gamma))
defaults = dict(lr=lr, betas=betas, final_lr=final_lr, gamma=gamma, eps=eps,
weight_decay=weight_decay, amsbound=amsbound)
super(AdaBound, self).__init__(params, defaults)
self.base_lrs = list(map(lambda group: group['lr'], self.param_groups))
def __setstate__(self, state):
super(AdaBound, self).__setstate__(state)
for group in self.param_groups:
group.setdefault('amsbound', False)
def step(self, closure=None):
"""Performs a single optimization step.
Arguments:
closure (callable, optional): A closure that reevaluates the model
and returns the loss.
"""
loss = None
if closure is not None:
loss = closure()
for group, base_lr in zip(self.param_groups, self.base_lrs):
for p in group['params']:
if p.grad is None:
continue
grad = p.grad.data
if grad.is_sparse:
raise RuntimeError(
'Adam does not support sparse gradients, please consider SparseAdam instead')
amsbound = group['amsbound']
state = self.state[p]
# State initialization
if len(state) == 0:
state['step'] = 0
# Exponential moving average of gradient values
state['exp_avg'] = torch.zeros_like(p.data)
# Exponential moving average of squared gradient values
state['exp_avg_sq'] = torch.zeros_like(p.data)
if amsbound:
# Maintains max of all exp. moving avg. of sq. grad. values
state['max_exp_avg_sq'] = torch.zeros_like(p.data)
exp_avg, exp_avg_sq = state['exp_avg'], state['exp_avg_sq']
if amsbound:
max_exp_avg_sq = state['max_exp_avg_sq']
beta1, beta2 = group['betas']
state['step'] += 1
if group['weight_decay'] != 0:
grad = grad.add(group['weight_decay'], p.data)
# Decay the first and second moment running average coefficient
exp_avg.mul_(beta1).add_(1 - beta1, grad)
exp_avg_sq.mul_(beta2).addcmul_(1 - beta2, grad, grad)
if amsbound:
# Maintains the maximum of all 2nd moment running avg. till now
torch.max(max_exp_avg_sq, exp_avg_sq, out=max_exp_avg_sq)
# Use the max. for normalizing running avg. of gradient
denom = max_exp_avg_sq.sqrt().add_(group['eps'])
else:
denom = exp_avg_sq.sqrt().add_(group['eps'])
bias_correction1 = 1 - beta1 ** state['step']
bias_correction2 = 1 - beta2 ** state['step']
step_size = group['lr'] * math.sqrt(bias_correction2) / bias_correction1
# Applies bounds on actual learning rate
# lr_scheduler cannot affect final_lr, this is a workaround to apply lr decay
final_lr = group['final_lr'] * group['lr'] / base_lr
lower_bound = final_lr * (1 - 1 / (group['gamma'] * state['step'] + 1))
upper_bound = final_lr * (1 + 1 / (group['gamma'] * state['step']))
step_size = torch.full_like(denom, step_size)
step_size.div_(denom).clamp_(lower_bound, upper_bound).mul_(exp_avg)
p.data.add_(-step_size)
return loss
class AdaBoundW(Optimizer):
"""Implements AdaBound algorithm with Decoupled Weight Decay (arxiv.org/abs/1711.05101)
It has been proposed in `Adaptive Gradient Methods with Dynamic Bound of Learning Rate`_.
Arguments:
params (iterable): iterable of parameters to optimize or dicts defining
parameter groups
lr (float, optional): Adam learning rate (default: 1e-3)
betas (Tuple[float, float], optional): coefficients used for computing
running averages of gradient and its square (default: (0.9, 0.999))
final_lr (float, optional): final (SGD) learning rate (default: 0.1)
gamma (float, optional): convergence speed of the bound functions (default: 1e-3)
eps (float, optional): term added to the denominator to improve
numerical stability (default: 1e-8)
weight_decay (float, optional): weight decay (L2 penalty) (default: 0)
amsbound (boolean, optional): whether to use the AMSBound variant of this algorithm
.. Adaptive Gradient Methods with Dynamic Bound of Learning Rate:
https://openreview.net/forum?id=Bkg3g2R9FX
"""
def __init__(self, params, lr=1e-3, betas=(0.9, 0.999), final_lr=0.1, gamma=1e-3,
eps=1e-8, weight_decay=0, amsbound=False):
if not 0.0 <= lr:
raise ValueError("Invalid learning rate: {}".format(lr))
if not 0.0 <= eps:
raise ValueError("Invalid epsilon value: {}".format(eps))
if not 0.0 <= betas[0] < 1.0:
raise ValueError("Invalid beta parameter at index 0: {}".format(betas[0]))
if not 0.0 <= betas[1] < 1.0:
raise ValueError("Invalid beta parameter at index 1: {}".format(betas[1]))
if not 0.0 <= final_lr:
raise ValueError("Invalid final learning rate: {}".format(final_lr))
if not 0.0 <= gamma < 1.0:
raise ValueError("Invalid gamma parameter: {}".format(gamma))
defaults = dict(lr=lr, betas=betas, final_lr=final_lr, gamma=gamma, eps=eps,
weight_decay=weight_decay, amsbound=amsbound)
super(AdaBoundW, self).__init__(params, defaults)
self.base_lrs = list(map(lambda group: group['lr'], self.param_groups))
def __setstate__(self, state):
super(AdaBoundW, self).__setstate__(state)
for group in self.param_groups:
group.setdefault('amsbound', False)
def step(self, closure=None):
"""Performs a single optimization step.
Arguments:
closure (callable, optional): A closure that reevaluates the model
and returns the loss.
"""
loss = None
if closure is not None:
loss = closure()
for group, base_lr in zip(self.param_groups, self.base_lrs):
for p in group['params']:
if p.grad is None:
continue
grad = p.grad.data
if grad.is_sparse:
raise RuntimeError(
'Adam does not support sparse gradients, please consider SparseAdam instead')
amsbound = group['amsbound']
state = self.state[p]
# State initialization
if len(state) == 0:
state['step'] = 0
# Exponential moving average of gradient values
state['exp_avg'] = torch.zeros_like(p.data)
# Exponential moving average of squared gradient values
state['exp_avg_sq'] = torch.zeros_like(p.data)
if amsbound:
# Maintains max of all exp. moving avg. of sq. grad. values
state['max_exp_avg_sq'] = torch.zeros_like(p.data)
exp_avg, exp_avg_sq = state['exp_avg'], state['exp_avg_sq']
if amsbound:
max_exp_avg_sq = state['max_exp_avg_sq']
beta1, beta2 = group['betas']
state['step'] += 1
# Decay the first and second moment running average coefficient
exp_avg.mul_(beta1).add_(1 - beta1, grad)
exp_avg_sq.mul_(beta2).addcmul_(1 - beta2, grad, grad)
if amsbound:
# Maintains the maximum of all 2nd moment running avg. till now
torch.max(max_exp_avg_sq, exp_avg_sq, out=max_exp_avg_sq)
# Use the max. for normalizing running avg. of gradient
denom = max_exp_avg_sq.sqrt().add_(group['eps'])
else:
denom = exp_avg_sq.sqrt().add_(group['eps'])
bias_correction1 = 1 - beta1 ** state['step']
bias_correction2 = 1 - beta2 ** state['step']
step_size = group['lr'] * math.sqrt(bias_correction2) / bias_correction1
# Applies bounds on actual learning rate
# lr_scheduler cannot affect final_lr, this is a workaround to apply lr decay
final_lr = group['final_lr'] * group['lr'] / base_lr
lower_bound = final_lr * (1 - 1 / (group['gamma'] * state['step'] + 1))
upper_bound = final_lr * (1 + 1 / (group['gamma'] * state['step']))
step_size = torch.full_like(denom, step_size)
step_size.div_(denom).clamp_(lower_bound, upper_bound).mul_(exp_avg)
if group['weight_decay'] != 0:
decayed_weights = torch.mul(p.data, group['weight_decay'])
p.data.add_(-step_size)
p.data.sub_(decayed_weights)
else:
p.data.add_(-step_size)
return loss