HairFastGAN

1
# Ranger deep learning optimizer - RAdam + Lookahead + Gradient Centralization, combined into one optimizer.
2

3
# https://github.com/lessw2020/Ranger-Deep-Learning-Optimizer
4
# and/or
5
# https://github.com/lessw2020/Best-Deep-Learning-Optimizers
6

7
# Ranger has now been used to capture 12 records on the FastAI leaderboard.
8

9
# This version = 20.4.11
10

11
# Credits:
12
# Gradient Centralization --> https://arxiv.org/abs/2004.01461v2 (a new optimization technique for DNNs), github:  https://github.com/Yonghongwei/Gradient-Centralization
13
# RAdam -->  https://github.com/LiyuanLucasLiu/RAdam
14
# Lookahead --> rewritten by lessw2020, but big thanks to Github @LonePatient and @RWightman for ideas from their code.
15
# Lookahead paper --> MZhang,G Hinton  https://arxiv.org/abs/1907.08610
16

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),
20
# 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;
22
# changed eps to 1e-5 as better default than 1e-8.
23

24
import math
25
import torch
26
from torch.optim.optimizer import Optimizer
27

28

29
class Ranger(Optimizer):
30

31
	def __init__(self, params, lr=1e-3,  # lr
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
				 use_gc=True, gc_conv_only=False
35
				 # Gradient centralization on or off, applied to conv layers only or conv + fc layers
36
				 ):
37

38
		# parameter checks
39
		if not 0.0 <= alpha <= 1.0:
40
			raise ValueError(f'Invalid slow update rate: {alpha}')
41
		if not 1 <= k:
42
			raise ValueError(f'Invalid lookahead steps: {k}')
43
		if not lr > 0:
44
			raise ValueError(f'Invalid Learning Rate: {lr}')
45
		if not eps > 0:
46
			raise ValueError(f'Invalid eps: {eps}')
47

48
		# parameter comments:
49
		# 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.
52

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,
55
						eps=eps, weight_decay=weight_decay)
56
		super().__init__(params, defaults)
57

58
		# adjustable threshold
59
		self.N_sma_threshhold = N_sma_threshhold
60

61
		# look ahead params
62

63
		self.alpha = alpha
64
		self.k = k
65

66
		# radam buffer for state
67
		self.radam_buffer = [[None, None, None] for ind in range(10)]
68

69
		# gc on or off
70
		self.use_gc = use_gc
71

72
		# level of gradient centralization
73
		self.gc_gradient_threshold = 3 if gc_conv_only else 1
74

75
	def __setstate__(self, state):
76
		super(Ranger, self).__setstate__(state)
77

78
	def step(self, closure=None):
79
		loss = None
80

81
		# Evaluate averages and grad, update param tensors
82
		for group in self.param_groups:
83

84
			for p in group['params']:
85
				if p.grad is None:
86
					continue
87
				grad = p.grad.data.float()
88

89
				if grad.is_sparse:
90
					raise RuntimeError('Ranger optimizer does not support sparse gradients')
91

92
				p_data_fp32 = p.data.float()
93

94
				state = self.state[p]  # get state dict for this param
95

96
				if len(state) == 0:  # if first time to run...init dictionary with our desired entries
97
					# if self.first_run_check==0:
98
					# self.first_run_check=1
99
					# print("Initializing slow buffer...should not see this at load from saved model!")
100
					state['step'] = 0
101
					state['exp_avg'] = torch.zeros_like(p_data_fp32)
102
					state['exp_avg_sq'] = torch.zeros_like(p_data_fp32)
103

104
					# look ahead weight storage now in state dict
105
					state['slow_buffer'] = torch.empty_like(p.data)
106
					state['slow_buffer'].copy_(p.data)
107

108
				else:
109
					state['exp_avg'] = state['exp_avg'].type_as(p_data_fp32)
110
					state['exp_avg_sq'] = state['exp_avg_sq'].type_as(p_data_fp32)
111

112
				# begin computations
113
				exp_avg, exp_avg_sq = state['exp_avg'], state['exp_avg_sq']
114
				beta1, beta2 = group['betas']
115

116
				# GC operation for Conv layers and FC layers
117
				if grad.dim() > self.gc_gradient_threshold:
118
					grad.add_(-grad.mean(dim=tuple(range(1, grad.dim())), keepdim=True))
119

120
				state['step'] += 1
121

122
				# compute variance mov avg
123
				exp_avg_sq.mul_(beta2).addcmul_(1 - beta2, grad, grad)
124
				# compute mean moving avg
125
				exp_avg.mul_(beta1).add_(1 - beta1, grad)
126

127
				buffered = self.radam_buffer[int(state['step'] % 10)]
128

129
				if state['step'] == buffered[0]:
130
					N_sma, step_size = buffered[1], buffered[2]
131
				else:
132
					buffered[0] = state['step']
133
					beta2_t = beta2 ** state['step']
134
					N_sma_max = 2 / (1 - beta2) - 1
135
					N_sma = N_sma_max - 2 * state['step'] * beta2_t / (1 - beta2_t)
136
					buffered[1] = N_sma
137
					if N_sma > self.N_sma_threshhold:
138
						step_size = math.sqrt(
139
							(1 - beta2_t) * (N_sma - 4) / (N_sma_max - 4) * (N_sma - 2) / N_sma * N_sma_max / (
140
										N_sma_max - 2)) / (1 - beta1 ** state['step'])
141
					else:
142
						step_size = 1.0 / (1 - beta1 ** state['step'])
143
					buffered[2] = step_size
144

145
				if group['weight_decay'] != 0:
146
					p_data_fp32.add_(-group['weight_decay'] * group['lr'], p_data_fp32)
147

148
				# apply lr
149
				if N_sma > self.N_sma_threshhold:
150
					denom = exp_avg_sq.sqrt().add_(group['eps'])
151
					p_data_fp32.addcdiv_(-step_size * group['lr'], exp_avg, denom)
152
				else:
153
					p_data_fp32.add_(-step_size * group['lr'], exp_avg)
154

155
				p.data.copy_(p_data_fp32)
156

157
				# integrated look ahead...
158
				# we do it at the param level instead of group level
159
				if state['step'] % group['k'] == 0:
160
					slow_p = state['slow_buffer']  # get access to slow param tensor
161
					slow_p.add_(self.alpha, p.data - slow_p)  # (fast weights - slow weights) * alpha
162
					p.data.copy_(slow_p)  # copy interpolated weights to RAdam param tensor
163

164
		return loss