Skip to content

Commit 1042b8a

Browse files
rwightmanrwightman
committed
Add non fused LAMB optimizer option
1 parent 01cb46a commit 1042b8a

File tree

2 files changed

+211
-1
lines changed

2 files changed

+211
-1
lines changed

‎timm/optim/lamb.py‎

Lines changed: 204 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,204 @@
1+
""" PyTorch Lamb optimizer w/ behaviour similar to NVIDIA FusedLamb
2+
3+
This optimizer code was adapted from the following (starting with latest)
4+
* https://github.com/HabanaAI/Model-References/blob/2b435114fe8e31f159b1d3063b8280ae37af7423/PyTorch/nlp/bert/pretraining/lamb.py
5+
* https://github.com/NVIDIA/DeepLearningExamples/blob/master/PyTorch/LanguageModeling/Transformer-XL/pytorch/lamb.py
6+
* https://github.com/cybertronai/pytorch-lamb
7+
8+
Use FusedLamb if you can. The reason for including this variant of Lamb is to have a version that is
9+
similar in behaviour to APEX FusedLamb if you aren't using NVIDIA GPUs or cannot install APEX for whatever reason.
10+
11+
Original copyrights for above sources are below.
12+
"""
13+
# Copyright (c) 2021, Habana Labs Ltd. All rights reserved.
14+
15+
# Copyright (c) 2019-2020, NVIDIA CORPORATION. All rights reserved.
16+
#
17+
# Licensed under the Apache License, Version 2.0 (the "License");
18+
# you may not use this file except in compliance with the License.
19+
# You may obtain a copy of the License at
20+
#
21+
# http://www.apache.org/licenses/LICENSE-2.0
22+
#
23+
# Unless required by applicable law or agreed to in writing, software
24+
# distributed under the License is distributed on an "AS IS" BASIS,
25+
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
26+
# See the License for the specific language governing permissions and
27+
# limitations under the License.
28+
29+
# MIT License
30+
#
31+
# Copyright (c) 2019 cybertronai
32+
#
33+
# Permission is hereby granted, free of charge, to any person obtaining a copy
34+
# of this software and associated documentation files (the "Software"), to deal
35+
# in the Software without restriction, including without limitation the rights
36+
# to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
37+
# copies of the Software, and to permit persons to whom the Software is
38+
# furnished to do so, subject to the following conditions:
39+
#
40+
# The above copyright notice and this permission notice shall be included in all
41+
# copies or substantial portions of the Software.
42+
#
43+
# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
44+
# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
45+
# FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
46+
# AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
47+
# LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
48+
# OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
49+
# SOFTWARE.
50+
51+
import torch
52+
from torch.optim import Optimizer
53+
54+
55+
class NvLamb(Optimizer):
56+
"""Implements a pure pytorch variant of FuseLAMB (NvLamb variant) optimizer from apex.optimizers.FusedLAMB
57+
reference: https://github.com/NVIDIA/DeepLearningExamples/blob/master/PyTorch/LanguageModeling/Transformer-XL/pytorch/lamb.py
58+
59+
LAMB was proposed in `Large Batch Optimization for Deep Learning: Training BERT in 76 minutes`_.
60+
61+
Arguments:
62+
params (iterable): iterable of parameters to optimize or dicts defining
63+
parameter groups.
64+
lr (float, optional): learning rate. (default: 1e-3)
65+
betas (Tuple[float, float], optional): coefficients used for computing
66+
running averages of gradient and its norm. (default: (0.9, 0.999))
67+
eps (float, optional): term added to the denominator to improve
68+
numerical stability. (default: 1e-8)
69+
weight_decay (float, optional): weight decay (L2 penalty) (default: 0)
70+
grad_averaging (bool, optional): whether apply (1-beta2) to grad when
71+
calculating running averages of gradient. (default: True)
72+
set_grad_none (bool, optional): whether set grad to None when zero_grad()
73+
method is called. (default: True)
74+
max_grad_norm (float, optional): value used to clip global grad norm
75+
(default: 1.0)
76+
use_nvlamb (boolean, optional): Apply adaptive learning rate to 0.0
77+
weight decay parameter (default: False)
78+
79+
.. _Large Batch Optimization for Deep Learning - Training BERT in 76 minutes:
80+
https://arxiv.org/abs/1904.00962
81+
.. _On the Convergence of Adam and Beyond:
82+
https://openreview.net/forum?id=ryQu7f-RZ
83+
"""
84+
85+
def __init__(self, params, lr=1e-3, bias_correction=True,
86+
betas=(0.9, 0.999), eps=1e-6, weight_decay=0.01,
87+
grad_averaging=True, set_grad_none=True,
88+
max_grad_norm=1.0, use_nvlamb=False):
89+
defaults = dict(lr=lr, bias_correction=bias_correction,
90+
betas=betas, eps=eps, weight_decay=weight_decay,
91+
grad_averaging=grad_averaging,
92+
max_grad_norm=max_grad_norm)
93+
super().__init__(params, defaults)
94+
self.set_grad_none = set_grad_none
95+
self.use_nvlamb = use_nvlamb
96+
97+
def zero_grad(self):
98+
if self.set_grad_none:
99+
for group in self.param_groups:
100+
for p in group['params']:
101+
p.grad = None
102+
else:
103+
super(NvLamb, self).zero_grad()
104+
105+
def step(self, closure=None):
106+
"""Performs a single optimization step.
107+
Arguments:
108+
closure (callable, optional): A closure that reevaluates the model
109+
and returns the loss.
110+
"""
111+
device = self.param_groups[0]["params"][0].device
112+
113+
loss = None
114+
if closure is not None:
115+
loss = closure()
116+
117+
global_grad_norm = torch.zeros(1, device=device)
118+
for group in self.param_groups:
119+
for p in group['params']:
120+
if p.grad is None:
121+
continue
122+
grad = p.grad.data
123+
if grad.is_sparse:
124+
raise RuntimeError('Lamb does not support sparse gradients, consider SparseAdam instad.')
125+
global_grad_norm.add_(grad.pow(2).sum())
126+
127+
global_grad_norm_ = torch.sqrt(global_grad_norm)
128+
max_grad_norm = self.defaults['max_grad_norm']
129+
130+
if global_grad_norm_ > max_grad_norm:
131+
clip_global_grad_norm = global_grad_norm_ / max_grad_norm
132+
else:
133+
clip_global_grad_norm = 1.0
134+
135+
for group in self.param_groups:
136+
bias_correction = 1 if group['bias_correction'] else 0
137+
beta1, beta2 = group['betas']
138+
grad_averaging = 1 if group['grad_averaging'] else 0
139+
if grad_averaging:
140+
beta3 = 1 - beta1
141+
else:
142+
beta3 = 1.0
143+
144+
# assume same step across group now to simplify things
145+
# per parameter step can be easily support by making it tensor, or pass list into kernel
146+
if 'step' in group:
147+
group['step'] += 1
148+
else:
149+
group['step'] = 1
150+
151+
step_size = group['lr']
152+
153+
if bias_correction:
154+
bias_correction1 = 1 - beta1 ** group['step']
155+
bias_correction2 = 1 - beta2 ** group['step']
156+
else:
157+
bias_correction1, bias_correction2 = 1.0, 1.0
158+
159+
for p in group['params']:
160+
if p.grad is None:
161+
continue
162+
grad = p.grad.data.div_(clip_global_grad_norm)
163+
state = self.state[p]
164+
165+
# State initialization
166+
if len(state) == 0:
167+
# Exponential moving average of gradient values
168+
state['exp_avg'] = torch.zeros_like(p.data)
169+
# Exponential moving average of squared gradient values
170+
state['exp_avg_sq'] = torch.zeros_like(p.data)
171+
172+
exp_avg_, exp_avg_sq_ = state['exp_avg'], state['exp_avg_sq']
173+
174+
# Decay the first and second moment running average coefficient
175+
# m_t
176+
exp_avg_.mul_(beta1).add_(grad, alpha=beta3)
177+
# v_t
178+
exp_avg_sq_.mul_(beta2).addcmul_(grad, grad, value=1 - beta2)
179+
# create clones to avoid modifying runner stats
180+
exp_avg = exp_avg_.div(bias_correction1)
181+
exp_avg_sq = exp_avg_sq_.div(bias_correction2)
182+
183+
# || w_t ||
184+
weight_norm = p.data.norm(2.0)
185+
# u_t
186+
exp_avg_sq_sqrt = torch.sqrt(exp_avg_sq)
187+
adam_step = exp_avg.div_(exp_avg_sq_sqrt.add_(group['eps']))
188+
if group['weight_decay'] != 0:
189+
adam_step.add_(p.data, alpha=group['weight_decay'])
190+
# || u_t ||
191+
adam_norm = adam_step.norm(2.0)
192+
if (group['weight_decay'] != 0 or self.use_nvlamb) and adam_norm > 0 and weight_norm > 0:
193+
trust_ratio = weight_norm / adam_norm
194+
trust_ratio = trust_ratio.item()
195+
else:
196+
trust_ratio = 1
197+
198+
state['weight_norm'] = weight_norm
199+
state['adam_norm'] = adam_norm
200+
state['trust_ratio'] = trust_ratio
201+
202+
p.data.add_(adam_step, alpha=-step_size * trust_ratio)
203+
204+
return loss

‎timm/optim/optim_factory.py‎

Lines changed: 7 additions & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -7,17 +7,18 @@
77
import torch.nn as nn
88
import torch.optim as optim
99

10+
from .adabelief import AdaBelief
1011
from .adafactor import Adafactor
1112
from .adahessian import Adahessian
1213
from .adamp import AdamP
14+
from .lamb import NvLamb
1315
from .lookahead import Lookahead
1416
from .nadam import Nadam
1517
from .novograd import NovoGrad
1618
from .nvnovograd import NvNovoGrad
1719
from .radam import RAdam
1820
from .rmsprop_tf import RMSpropTF
1921
from .sgdp import SGDP
20-
from .adabelief import AdaBelief
2122

2223
try:
2324
from apex.optimizers import FusedNovoGrad, FusedAdam, FusedLAMB, FusedSGD
@@ -148,6 +149,10 @@ def create_optimizer_v2(
148149
optimizer = NovoGrad(parameters, **opt_args)
149150
elif opt_lower == 'nvnovograd':
150151
optimizer = NvNovoGrad(parameters, **opt_args)
152+
elif opt_lower == 'lamb':
153+
optimizer = NvLamb(parameters, **opt_args)
154+
155+
# NVIDIA fused optimizers, require APEX to be installed
151156
elif opt_lower == 'fusedsgd':
152157
opt_args.pop('eps', None)
153158
optimizer = FusedSGD(parameters, momentum=momentum, nesterov=True, **opt_args)
@@ -163,6 +168,7 @@ def create_optimizer_v2(
163168
elif opt_lower == 'fusednovograd':
164169
opt_args.setdefault('betas', (0.95, 0.98))
165170
optimizer = FusedNovoGrad(parameters, **opt_args)
171+
166172
else:
167173
assert False and "Invalid optimizer"
168174
raise ValueError

0 commit comments

Comments
 (0)