Skip to content

Commit 81d386b

Browse files
lessw2020lessw2020
authored
formatted with pylint
1 parent 8d636a5 commit 81d386b

File tree

1 file changed

+72
-76
lines changed

1 file changed

+72
-76
lines changed

‎ranger/ranger.py‎

Lines changed: 72 additions & 76 deletions
Original file line numberDiff line numberDiff line change
@@ -1,41 +1,41 @@
11
# Ranger deep learning optimizer - RAdam + Lookahead + Gradient Centralization, combined into one optimizer.
22

3-
# https://github.com/lessw2020/Ranger-Deep-Learning-Optimizer
3+
# https://github.com/lessw2020/Ranger-Deep-Learning-Optimizer
44
# and/or
55
# https://github.com/lessw2020/Best-Deep-Learning-Optimizers
66

77
# Ranger has now been used to capture 12 records on the FastAI leaderboard.
88

9-
# This version = 20.4.11
9+
# This version = 20.4.11
1010

1111
# Credits:
1212
# Gradient Centralization --> https://arxiv.org/abs/2004.01461v2 (a new optimization technique for DNNs), github: https://github.com/Yonghongwei/Gradient-Centralization
1313
# RAdam --> https://github.com/LiyuanLucasLiu/RAdam
1414
# Lookahead --> rewritten by lessw2020, but big thanks to Github @LonePatient and @RWightman for ideas from their code.
1515
# Lookahead paper --> MZhang,G Hinton https://arxiv.org/abs/1907.08610
1616

17-
# summary of changes:
18-
# 4/11/20 - add gradient centralization option. Set new testing benchmark for accuracy with it, toggle with use_gc flag at init.
19-
# full code integration with all updates at param level instead of group, moves slow weights into state dict (from generic weights),
17+
# summary of changes:
18+
# 4/11/20 - add gradient centralization option. Set new testing benchmark for accuracy with it, toggle with use_gc flag at init.
19+
# full code integration with all updates at param level instead of group, moves slow weights into state dict (from generic weights),
2020
# supports group learning rates (thanks @SHolderbach), fixes sporadic load from saved model issues.
21-
# changes 8/31/19 - fix references to *self*.N_sma_threshold;
21+
# changes 8/31/19 - fix references to *self*.N_sma_threshold;
2222
# changed eps to 1e-5 as better default than 1e-8.
2323

2424
import math
2525
import torch
2626
from torch.optim.optimizer import Optimizer, required
2727

2828

29-
3029
class Ranger(Optimizer):
3130

3231
def __init__(self, params, lr=1e-3, # lr
33-
alpha=0.5, k=6, N_sma_threshhold=5, # Ranger options
34-
betas=(.95,0.999), eps=1e-5, weight_decay=0, # Adam options
35-
use_gc=True, gc_conv_only=False # Gradient centralization on or off, applied to conv layers only or conv + fc layers
36-
):
32+
alpha=0.5, k=6, N_sma_threshhold=5, # Ranger options
33+
betas=(.95, 0.999), eps=1e-5, weight_decay=0, # Adam options
34+
# Gradient centralization on or off, applied to conv layers only or conv + fc layers
35+
use_gc=True, gc_conv_only=False
36+
):
3737

38-
#parameter checks
38+
# parameter checks
3939
if not 0.0 <= alpha <= 1.0:
4040
raise ValueError(f'Invalid slow update rate: {alpha}')
4141
if not 1 <= k:
@@ -45,58 +45,53 @@ def __init__(self, params, lr=1e-3, # lr
4545
if not eps > 0:
4646
raise ValueError(f'Invalid eps: {eps}')
4747

48-
#parameter comments:
48+
# parameter comments:
4949
# beta1 (momentum) of .95 seems to work better than .90...
50-
#N_sma_threshold of 5 seems better in testing than 4.
51-
#In both cases, worth testing on your dataset (.90 vs .95, 4 vs 5) to make sure which works best for you.
50+
# N_sma_threshold of 5 seems better in testing than 4.
51+
# In both cases, worth testing on your dataset (.90 vs .95, 4 vs 5) to make sure which works best for you.
5252

53-
#prep defaults and init torch.optim base
54-
defaults = dict(lr=lr, alpha=alpha, k=k, step_counter=0, betas=betas, N_sma_threshhold=N_sma_threshhold, eps=eps, weight_decay=weight_decay)
55-
super().__init__(params,defaults)
53+
# prep defaults and init torch.optim base
54+
defaults = dict(lr=lr, alpha=alpha, k=k, step_counter=0, betas=betas,
55+
N_sma_threshhold=N_sma_threshhold, eps=eps, weight_decay=weight_decay)
56+
super().__init__(params, defaults)
5657

57-
#adjustable threshold
58+
# adjustable threshold
5859
self.N_sma_threshhold = N_sma_threshhold
5960

60-
61-
#look ahead params
61+
# look ahead params
6262

6363
self.alpha = alpha
64-
self.k = k
64+
self.k = k
65+
66+
# radam buffer for state
67+
self.radam_buffer = [[None, None, None] for ind in range(10)]
6568

66-
#radam buffer for state
67-
self.radam_buffer = [[None,None,None] for ind in range(10)]
69+
# gc on or off
70+
self.use_gc = use_gc
6871

69-
#gc on or off
70-
self.use_gc=use_gc
71-
72-
#level of gradient centralization
72+
# level of gradient centralization
7373
self.gc_gradient_threshold = 3 if gc_conv_only else 1
74-
75-
76-
print(f"Ranger optimizer loaded. \nGradient Centralization usage = {self.use_gc}")
77-
if (self.use_gc and self.gc_gradient_threshold==1):
74+
75+
print(
76+
f"Ranger optimizer loaded. \nGradient Centralization usage = {self.use_gc}")
77+
if (self.use_gc and self.gc_gradient_threshold == 1):
7878
print(f"GC applied to both conv and fc layers")
79-
elif (self.use_gc and self.gc_gradient_threshold==3):
79+
elif (self.use_gc and self.gc_gradient_threshold == 3):
8080
print(f"GC applied to conv layers only")
81-
82-
83-
84-
8581

8682
def __setstate__(self, state):
8783
print("set state called")
8884
super(Ranger, self).__setstate__(state)
8985

90-
9186
def step(self, closure=None):
9287
loss = None
93-
#note - below is commented out b/c I have other work that passes back the loss as a float, and thus not a callable closure.
94-
#Uncomment if you need to use the actual closure...
88+
# note - below is commented out b/c I have other work that passes back the loss as a float, and thus not a callable closure.
89+
# Uncomment if you need to use the actual closure...
9590

96-
#if closure is not None:
97-
#loss = closure()
91+
# if closure is not None:
92+
#loss = closure()
9893

99-
#Evaluate averages and grad, update param tensors
94+
# Evaluate averages and grad, update param tensors
10095
for group in self.param_groups:
10196

10297
for p in group['params']:
@@ -105,84 +100,85 @@ def step(self, closure=None):
105100
grad = p.grad.data.float()
106101

107102
if grad.is_sparse:
108-
raise RuntimeError('Ranger optimizer does not support sparse gradients')
103+
raise RuntimeError(
104+
'Ranger optimizer does not support sparse gradients')
109105

110106
p_data_fp32 = p.data.float()
111107

112-
state = self.state[p] #get state dict for this param
108+
state = self.state[p] # get state dict for this param
113109

114-
if len(state) == 0: #if first time to run...init dictionary with our desired entries
115-
#if self.first_run_check==0:
116-
#self.first_run_check=1
117-
#print("Initializing slow buffer...should not see this at load from saved model!")
110+
if len(state) == 0: # if first time to run...init dictionary with our desired entries
111+
# if self.first_run_check==0:
112+
# self.first_run_check=1
113+
#print("Initializing slow buffer...should not see this at load from saved model!")
118114
state['step'] = 0
119115
state['exp_avg'] = torch.zeros_like(p_data_fp32)
120116
state['exp_avg_sq'] = torch.zeros_like(p_data_fp32)
121117

122-
#look ahead weight storage now in state dict
118+
# look ahead weight storage now in state dict
123119
state['slow_buffer'] = torch.empty_like(p.data)
124120
state['slow_buffer'].copy_(p.data)
125121

126122
else:
127123
state['exp_avg'] = state['exp_avg'].type_as(p_data_fp32)
128-
state['exp_avg_sq'] = state['exp_avg_sq'].type_as(p_data_fp32)
124+
state['exp_avg_sq'] = state['exp_avg_sq'].type_as(
125+
p_data_fp32)
129126

130-
#begin computations
127+
# begin computations
131128
exp_avg, exp_avg_sq = state['exp_avg'], state['exp_avg_sq']
132129
beta1, beta2 = group['betas']
133-
134-
135-
#GC operation for Conv layers and FC layers
136-
if grad.dim() > self.gc_gradient_threshold:
137-
grad.add_(-grad.mean(dim = tuple(range(1,grad.dim())), keepdim = True))
138-
139-
130+
131+
# GC operation for Conv layers and FC layers
132+
if grad.dim() > self.gc_gradient_threshold:
133+
grad.add_(-grad.mean(dim=tuple(range(1, grad.dim())), keepdim=True))
140134

141135
state['step'] += 1
142136

143-
#compute variance mov avg
137+
# compute variance mov avg
144138
exp_avg_sq.mul_(beta2).addcmul_(1 - beta2, grad, grad)
145-
#compute mean moving avg
139+
# compute mean moving avg
146140
exp_avg.mul_(beta1).add_(1 - beta1, grad)
147141

148-
149-
150-
151-
152142
buffered = self.radam_buffer[int(state['step'] % 10)]
153-
143+
154144
if state['step'] == buffered[0]:
155145
N_sma, step_size = buffered[1], buffered[2]
156146
else:
157147
buffered[0] = state['step']
158148
beta2_t = beta2 ** state['step']
159149
N_sma_max = 2 / (1 - beta2) - 1
160-
N_sma = N_sma_max - 2 * state['step'] * beta2_t / (1 - beta2_t)
150+
N_sma = N_sma_max - 2 * \
151+
state['step'] * beta2_t / (1 - beta2_t)
161152
buffered[1] = N_sma
162153
if N_sma > self.N_sma_threshhold:
163-
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'])
154+
step_size = math.sqrt((1 - beta2_t) * (N_sma - 4) / (N_sma_max - 4) * (
155+
N_sma - 2) / N_sma * N_sma_max / (N_sma_max - 2)) / (1 - beta1 ** state['step'])
164156
else:
165157
step_size = 1.0 / (1 - beta1 ** state['step'])
166158
buffered[2] = step_size
167159

168-
169160
if group['weight_decay'] != 0:
170-
p_data_fp32.add_(-group['weight_decay'] * group['lr'], p_data_fp32)
161+
p_data_fp32.add_(-group['weight_decay']
162+
* group['lr'], p_data_fp32)
171163

172164
# apply lr
173165
if N_sma > self.N_sma_threshhold:
174166
denom = exp_avg_sq.sqrt().add_(group['eps'])
175-
p_data_fp32.addcdiv_(-step_size * group['lr'], exp_avg, denom)
167+
p_data_fp32.addcdiv_(-step_size *
168+
group['lr'], exp_avg, denom)
176169
else:
177170
p_data_fp32.add_(-step_size * group['lr'], exp_avg)
178171

179172
p.data.copy_(p_data_fp32)
180173

181-
#integrated look ahead...
182-
#we do it at the param level instead of group level
174+
# integrated look ahead...
175+
# we do it at the param level instead of group level
183176
if state['step'] % group['k'] == 0:
184-
slow_p = state['slow_buffer'] #get access to slow param tensor
185-
slow_p.add_(self.alpha, p.data - slow_p) #(fast weights - slow weights) * alpha
186-
p.data.copy_(slow_p) #copy interpolated weights to RAdam param tensor
177+
# get access to slow param tensor
178+
slow_p = state['slow_buffer']
179+
# (fast weights - slow weights) * alpha
180+
slow_p.add_(self.alpha, p.data - slow_p)
181+
# copy interpolated weights to RAdam param tensor
182+
p.data.copy_(slow_p)
187183

188184
return loss

0 commit comments

Comments
 (0)