File size: 6,465 Bytes
ecf08bc |
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 |
############
# https://github.com/lessw2020/Ranger-Deep-Learning-Optimizer
# This code was taken from the repo above and was not created by me (Fabian)! Full credit goes to the original authors
############
import math
import torch
from torch.optim.optimizer import Optimizer
class Ranger(Optimizer):
def __init__(self, params, lr=1e-3, alpha=0.5, k=6, N_sma_threshhold=5, betas=(.95, 0.999), eps=1e-5,
weight_decay=0):
# 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_threshhold=N_sma_threshhold,
eps=eps, weight_decay=weight_decay)
super().__init__(params, defaults)
# adjustable threshold
self.N_sma_threshhold = N_sma_threshhold
# now we can get to work...
# removed as we now use step from RAdam...no need for duplicate step counting
# for group in self.param_groups:
# group["step_counter"] = 0
# print("group step counter init")
# look ahead params
self.alpha = alpha
self.k = k
# radam buffer for state
self.radam_buffer = [[None, None, None] for ind in range(10)]
# self.first_run_check=0
# lookahead weights
# 9/2/19 - lookahead param tensors have been moved to state storage.
# This should resolve issues with load/save where weights were left in GPU memory from first load, slowing down future runs.
# self.slow_weights = [[p.clone().detach() for p in group['params']]
# for group in self.param_groups]
# don't use grad for lookahead weights
# for w in it.chain(*self.slow_weights):
# w.requires_grad = False
def __setstate__(self, state):
print("set state called")
super(Ranger, self).__setstate__(state)
def step(self, closure=None):
loss = None
# 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.
# Uncomment if you need to use the actual closure...
# if closure is not None:
# 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']
# compute variance mov avg
exp_avg_sq.mul_(beta2).addcmul_(1 - beta2, grad, grad)
# compute mean moving avg
exp_avg.mul_(beta1).add_(1 - beta1, grad)
state['step'] += 1
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_threshhold:
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)
if N_sma > self.N_sma_threshhold:
denom = exp_avg_sq.sqrt().add_(group['eps'])
p_data_fp32.addcdiv_(-step_size * group['lr'], exp_avg, denom)
else:
p_data_fp32.add_(-step_size * group['lr'], exp_avg)
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:
slow_p = state['slow_buffer'] # get access to slow param tensor
slow_p.add_(self.alpha, p.data - slow_p) # (fast weights - slow weights) * alpha
p.data.copy_(slow_p) # copy interpolated weights to RAdam param tensor
return loss
|