-
-
Notifications
You must be signed in to change notification settings - Fork 18
/
ranger.py
205 lines (161 loc) · 8.72 KB
/
ranger.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
# Ranger deep learning optimizer - RAdam + Lookahead + Gradient Centralization, combined into one optimizer.
# https://github.com/lessw2020/Ranger-Deep-Learning-Optimizer
# and/or
# https://github.com/lessw2020/Best-Deep-Learning-Optimizers
# Ranger has been used to capture 12 records on the FastAI leaderboard.
# This version = 2020.9.4
# Credits:
# Gradient Centralization --> https://arxiv.org/abs/2004.01461v2 (a new optimization technique for DNNs), github: https://github.com/Yonghongwei/Gradient-Centralization
# RAdam --> https://github.com/LiyuanLucasLiu/RAdam
# Lookahead --> rewritten by lessw2020, but big thanks to Github @LonePatient and @RWightman for ideas from their code.
# Lookahead paper --> MZhang,G Hinton https://arxiv.org/abs/1907.08610
# summary of changes:
# 9/4/20 - updated addcmul_ signature to avoid warning. Integrates latest changes from GC developer (he did the work for this), and verified on performance on private dataset.
# 4/11/20 - add gradient centralization option. Set new testing benchmark for accuracy with it, toggle with use_gc flag at init.
# full code integration with all updates at param level instead of group, moves slow weights into state dict (from generic weights),
# supports group learning rates (thanks @SHolderbach), fixes sporadic load from saved model issues.
# changes 8/31/19 - fix references to *self*.N_sma_threshold;
# changed eps to 1e-5 as better default than 1e-8.
import math
import torch
from torch.optim.optimizer import Optimizer, required
# If dim is None it will be chosen automatically
def centralized_gradient(x, use_gc=True, gc_conv_only=False, dim=None):
'''credit - https://github.com/Yonghongwei/Gradient-Centralization '''
if use_gc:
dim_threshold = 3 if gc_conv_only else 1
if len(list(x.size())) > dim_threshold:
x.add_(-x.mean(dim=(dim or tuple(range(1, len(list(x.size()))))), keepdim=True))
return x
class Ranger(Optimizer):
def __init__(self, params, lr=1e-3, # lr
alpha=0.5, k=6, N_sma_threshold=5, # Ranger options
betas=(.95, 0.999), eps=1e-5, weight_decay=0, # Adam options
# Gradient centralization on or off, applied to conv layers only or conv + fc layers
use_gc=True, gc_conv_only=False, gc_loc=True
):
# parameter checks
if not 0.0 <= alpha <= 1.0:
raise ValueError(f'Invalid slow update rate: {alpha}')
if not 1 <= k:
raise ValueError(f'Invalid lookahead steps: {k}')
if not lr > 0:
raise ValueError(f'Invalid Learning Rate: {lr}')
if not eps > 0:
raise ValueError(f'Invalid eps: {eps}')
# parameter comments:
# beta1 (momentum) of .95 seems to work better than .90...
# N_sma_threshold of 5 seems better in testing than 4.
# In both cases, worth testing on your dataset (.90 vs .95, 4 vs 5) to make sure which works best for you.
# prep defaults and init torch.optim base
defaults = dict(lr=lr, alpha=alpha, k=k, step_counter=0, betas=betas,
N_sma_threshold=N_sma_threshold, eps=eps, weight_decay=weight_decay,
gc_dim=None)
super().__init__(params, defaults)
# adjustable threshold
self.N_sma_threshold = N_sma_threshold
# look ahead params
self.alpha = alpha
self.k = k
# radam buffer for state
self.radam_buffer = [[None, None, None] for ind in range(10)]
# gc on or off
self.gc_loc = gc_loc
self.use_gc = use_gc
self.gc_conv_only = gc_conv_only
# level of gradient centralization
#self.gc_gradient_threshold = 3 if gc_conv_only else 1
print(
f"Ranger optimizer loaded. \nGradient Centralization usage = {self.use_gc}")
if (self.use_gc and self.gc_conv_only == False):
print(f"GC applied to both conv and fc layers")
elif (self.use_gc and self.gc_conv_only == True):
print(f"GC applied to conv layers only")
def __setstate__(self, state):
print("set state called")
super(Ranger, self).__setstate__(state)
def step(self, closure=None):
loss = None
if closure is not None:
with torch.enable_grad():
loss = closure()
# Evaluate averages and grad, update param tensors
for group in self.param_groups:
for p in group['params']:
if p.grad is None:
continue
grad = p.grad.data.float()
if grad.is_sparse:
raise RuntimeError(
'Ranger optimizer does not support sparse gradients')
p_data_fp32 = p.data.float()
state = self.state[p] # get state dict for this param
if len(state) == 0: # if first time to run...init dictionary with our desired entries
# if self.first_run_check==0:
# self.first_run_check=1
#print("Initializing slow buffer...should not see this at load from saved model!")
state['step'] = 0
state['exp_avg'] = torch.zeros_like(p_data_fp32)
state['exp_avg_sq'] = torch.zeros_like(p_data_fp32)
# look ahead weight storage now in state dict
state['slow_buffer'] = torch.empty_like(p.data)
state['slow_buffer'].copy_(p.data)
else:
state['exp_avg'] = state['exp_avg'].type_as(p_data_fp32)
state['exp_avg_sq'] = state['exp_avg_sq'].type_as(
p_data_fp32)
# begin computations
exp_avg, exp_avg_sq = state['exp_avg'], state['exp_avg_sq']
beta1, beta2 = group['betas']
# GC operation for Conv layers and FC layers
# if grad.dim() > self.gc_gradient_threshold:
# grad.add_(-grad.mean(dim=tuple(range(1, grad.dim())), keepdim=True))
if self.gc_loc:
grad = centralized_gradient(grad, use_gc=self.use_gc, gc_conv_only=self.gc_conv_only, dim=group['gc_dim'])
state['step'] += 1
# compute variance mov avg
exp_avg_sq.mul_(beta2).addcmul_(grad, grad, value=1 - beta2)
# compute mean moving avg
exp_avg.mul_(beta1).add_(grad, alpha=1 - beta1)
buffered = self.radam_buffer[int(state['step'] % 10)]
if state['step'] == buffered[0]:
N_sma, step_size = buffered[1], buffered[2]
else:
buffered[0] = state['step']
beta2_t = beta2 ** state['step']
N_sma_max = 2 / (1 - beta2) - 1
N_sma = N_sma_max - 2 * \
state['step'] * beta2_t / (1 - beta2_t)
buffered[1] = N_sma
if N_sma > self.N_sma_threshold:
step_size = math.sqrt((1 - beta2_t) * (N_sma - 4) / (N_sma_max - 4) * (
N_sma - 2) / N_sma * N_sma_max / (N_sma_max - 2)) / (1 - beta1 ** state['step'])
else:
step_size = 1.0 / (1 - beta1 ** state['step'])
buffered[2] = step_size
# if group['weight_decay'] != 0:
# p_data_fp32.add_(-group['weight_decay']
# * group['lr'], p_data_fp32)
# apply lr
if N_sma > self.N_sma_threshold:
denom = exp_avg_sq.sqrt().add_(group['eps'])
G_grad = exp_avg / denom
else:
G_grad = exp_avg
if group['weight_decay'] != 0:
G_grad.add_(p_data_fp32, alpha=group['weight_decay'])
# GC operation
if self.gc_loc == False:
G_grad = centralized_gradient(G_grad, use_gc=self.use_gc, gc_conv_only=self.gc_conv_only, dim=group['gc_dim'])
p_data_fp32.add_(G_grad, alpha=-step_size * group['lr'])
p.data.copy_(p_data_fp32)
# integrated look ahead...
# we do it at the param level instead of group level
if state['step'] % group['k'] == 0:
# get access to slow param tensor
slow_p = state['slow_buffer']
# (fast weights - slow weights) * alpha
slow_p.add_(p.data - slow_p, alpha=self.alpha)
# copy interpolated weights to RAdam param tensor
p.data.copy_(slow_p)
return loss