# 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
[docs]
def centralized_gradient(x, use_gc=True, gc_conv_only=False):
"""credit - https://github.com/Yonghongwei/Gradient-Centralization"""
if use_gc:
if gc_conv_only:
if len(list(x.size())) > 3:
x.add_(-x.mean(dim=tuple(range(1, len(list(x.size())))), keepdim=True))
else:
if len(list(x.size())) > 1:
x.add_(-x.mean(dim=tuple(range(1, len(list(x.size())))), keepdim=True))
return x
[docs]
class Ranger(Optimizer):
[docs]
def __init__(
self,
params,
lr=1e-3, # lr
alpha=0.5,
k=6,
N_sma_threshhold=5, # Ranger options
betas=(0.95, 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_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.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)
[docs]
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"]
# 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
)
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_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"])
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
)
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