-
Notifications
You must be signed in to change notification settings - Fork 26
/
optimization.py
165 lines (138 loc) · 6.6 KB
/
optimization.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
# coding=utf-8
# Copyright 2018 The Google AI Language Team Authors and The HugginFace Inc. team.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
"""PyTorch optimization for BERT model."""
import math
import torch
from torch.optim import Optimizer
from torch.nn.utils import clip_grad_norm_
def warmup_cosine(x, warmup=0.002):
if x < warmup:
return x / warmup
return 0.5 * (1.0 + torch.cos(math.pi * x))
def warmup_constant(x, warmup=0.002):
if x < warmup:
return x / warmup
return 1.0
def warmup_linear(x, warmup=0.002):
if x < warmup:
return x / warmup
return 1.0 - x
SCHEDULES = {
'warmup_cosine': warmup_cosine,
'warmup_constant': warmup_constant,
'warmup_linear': warmup_linear,
}
class BERTAdam(Optimizer):
"""Implements BERT version of Adam algorithm with weight decay fix (and no ).
Params:
lr: learning rate
warmup: portion of t_total for the warmup, -1 means no warmup. Default: -1
t_total: total number of training steps for the learning
rate schedule, -1 means constant learning rate. Default: -1
schedule: schedule to use for the warmup (see above). Default: 'warmup_linear'
b1: Adams b1. Default: 0.9
b2: Adams b2. Default: 0.999
e: Adams epsilon. Default: 1e-6
weight_decay_rate: Weight decay. Default: 0.01
max_grad_norm: Maximum norm for the gradients (-1 means no clipping). Default: 1.0
"""
def __init__(self, params, lr, warmup=-1, t_total=-1, schedule='warmup_linear',
b1=0.9, b2=0.999, e=1e-6, weight_decay_rate=0.01,
max_grad_norm=1.0):
if not lr >= 0.0:
raise ValueError("Invalid learning rate: {} - should be >= 0.0".format(lr))
if schedule not in SCHEDULES:
raise ValueError("Invalid schedule parameter: {}".format(schedule))
if not 0.0 <= warmup < 1.0 and not warmup == -1:
raise ValueError("Invalid warmup: {} - should be in [0.0, 1.0[ or -1".format(warmup))
if not 0.0 <= b1 < 1.0:
raise ValueError("Invalid b1 parameter: {} - should be in [0.0, 1.0[".format(b1))
if not 0.0 <= b2 < 1.0:
raise ValueError("Invalid b2 parameter: {} - should be in [0.0, 1.0[".format(b2))
if not e >= 0.0:
raise ValueError("Invalid epsilon value: {} - should be >= 0.0".format(e))
defaults = dict(lr=lr, schedule=schedule, warmup=warmup, t_total=t_total,
b1=b1, b2=b2, e=e, weight_decay_rate=weight_decay_rate,
max_grad_norm=max_grad_norm)
super(BERTAdam, self).__init__(params, defaults)
def get_lr(self):
lr = []
for group in self.param_groups:
for p in group['params']:
state = self.state[p]
if len(state) == 0:
return [0]
if group['t_total'] != -1:
schedule_fct = SCHEDULES[group['schedule']]
lr_scheduled = group['lr'] * schedule_fct(state['step'] / group['t_total'], group['warmup'])
else:
lr_scheduled = group['lr']
lr.append(lr_scheduled)
return lr
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 in self.param_groups:
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')
state = self.state[p]
# State initialization
if len(state) == 0:
state['step'] = 0
# Exponential moving average of gradient values
state['next_m'] = torch.zeros_like(p.data)
# Exponential moving average of squared gradient values
state['next_v'] = torch.zeros_like(p.data)
next_m, next_v = state['next_m'], state['next_v']
beta1, beta2 = group['b1'], group['b2']
# Add grad clipping
if group['max_grad_norm'] > 0:
clip_grad_norm_(p, group['max_grad_norm'])
# Decay the first and second moment running average coefficient
# In-place operations to update the averages at the same time
next_m.mul_(beta1).add_(1 - beta1, grad)
next_v.mul_(beta2).addcmul_(1 - beta2, grad, grad)
update = next_m / (next_v.sqrt() + group['e'])
# Just adding the square of the weights to the loss function is *not*
# the correct way of using L2 regularization/weight decay with Adam,
# since that will interact with the m and v parameters in strange ways.
#
# Instead we want ot decay the weights in a manner that doesn't interact
# with the m/v parameters. This is equivalent to adding the square
# of the weights to the loss with plain (non-momentum) SGD.
if group['weight_decay_rate'] > 0.0:
update += group['weight_decay_rate'] * p.data
if group['t_total'] != -1:
schedule_fct = SCHEDULES[group['schedule']]
lr_scheduled = group['lr'] * schedule_fct(state['step'] / group['t_total'], group['warmup'])
else:
lr_scheduled = group['lr']
update_with_lr = lr_scheduled * update
p.data.add_(-update_with_lr)
state['step'] += 1
# step_size = lr_scheduled * math.sqrt(bias_correction2) / bias_correction1
# bias_correction1 = 1 - beta1 ** state['step']
# bias_correction2 = 1 - beta2 ** state['step']
return loss