adirik's picture
update repo
f4c3c2b
# 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 now been used to capture 12 records on the FastAI leaderboard.
# This version = 20.4.11
# 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:
# 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
class Ranger(Optimizer):
def __init__(self, params, lr=1e-3, # lr
alpha=0.5, k=6, N_sma_threshhold=5, # Ranger options
betas=(.95, 0.999), eps=1e-5, weight_decay=0, # Adam options
use_gc=True, gc_conv_only=False
# Gradient centralization on or off, applied to conv layers only or conv + fc layers
):
# 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
# 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.use_gc = use_gc
# level of gradient centralization
self.gc_gradient_threshold = 3 if gc_conv_only else 1
def __setstate__(self, state):
super(Ranger, self).__setstate__(state)
def step(self, closure=None):
loss = None
# 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))
state['step'] += 1
# 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)
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)
# apply lr
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