10/19 08:14

AdaBelief

• 论文链接：https://arxiv.org/pdf/2010.07468.pdf

import math
import torch
from torch.optim.optimizer import Optimizer

version_higher = ( torch.__version__ >= "1.5.0" )

Arguments:
params (iterable): iterable of parameters to optimize or dicts defining
parameter groups
lr (float, optional): learning rate (default: 1e-3)
betas (Tuple[float, float], optional): coefficients used for computing
running averages of gradient and its square (default: (0.9, 0.999))
eps (float, optional): term added to the denominator to improve
numerical stability (default: 1e-8)
weight_decay (float, optional): weight decay (L2 penalty) (default: 0)
algorithm from the paper On the Convergence of Adam and Beyond_
(default: False)
weight_decouple (boolean, optional): ( default: False) If set as True, then
the optimizer uses decoupled weight decay as in AdamW
fixed_decay (boolean, optional): (default: False) This is used when weight_decouple
is set as True.
When fixed_decay == True, the weight decay is performed as
$W_{new} = W_{old} - W_{old} \times decay$.
When fixed_decay == False, the weight decay is performed as
$W_{new} = W_{old} - W_{old} \times decay \times lr$. Note that in this case, the
weight decay ratio decreases with learning rate (lr).
rectify (boolean, optional): (default: False) If set as True, then perform the rectified

NeurIPS 2020 Spotlight
"""

def __init__(self, params, lr=1e-3, betas=(0.9, 0.999), eps=1e-8,
weight_decay=0, amsgrad=False, weight_decouple = False, fixed_decay=False, rectify = False ):
if not 0.0 <= lr:
raise ValueError("Invalid learning rate: {}".format(lr))
if not 0.0 <= eps:
raise ValueError("Invalid epsilon value: {}".format(eps))
if not 0.0 <= betas[0] < 1.0:
raise ValueError("Invalid beta parameter at index 0: {}".format(betas[0]))
if not 0.0 <= betas[1] < 1.0:
raise ValueError("Invalid beta parameter at index 1: {}".format(betas[1]))
defaults = dict(lr=lr, betas=betas, eps=eps,

self.weight_decouple = weight_decouple
self.rectify = rectify
self.fixed_decay = fixed_decay
if self.weight_decouple:
if self.fixed_decay:
print('Weight decay fixed')
if self.rectify:
def __setstate__(self, state):
for group in self.param_groups:

def reset(self):
for group in self.param_groups:
for p in group['params']:
state = self.state[p]

# State initialization
state['step'] = 0
# Exponential moving average of gradient values
state['exp_avg'] = torch.zeros_like(p.data,
memory_format=torch.preserve_format) if version_higher else torch.zeros_like(p.data)

# Exponential moving average of squared gradient values
state['exp_avg_var'] = torch.zeros_like(p.data,
memory_format=torch.preserve_format) if version_higher else torch.zeros_like(p.data)
# Maintains max of all exp. moving avg. of sq. grad. values
state['max_exp_avg_var'] = torch.zeros_like(p.data,
memory_format=torch.preserve_format) if version_higher else torch.zeros_like(p.data)

def step(self, closure=None):
"""Performs a single optimization step.

Arguments:
closure (callable, optional): A closure that reevaluates the model
and returns the loss.
"""
loss = None
if closure is not None:
loss = closure()

for group in self.param_groups:
for p in group['params']:
continue

state = self.state[p]

beta1, beta2 = group['betas']

# State initialization
if len(state) == 0:
state['rho_inf'] = 2.0 / (1.0 - beta2) - 1.0
state['step'] = 0
# Exponential moving average of gradient values
state['exp_avg'] = torch.zeros_like(p.data,
memory_format=torch.preserve_format) if version_higher else torch.zeros_like(p.data)
# Exponential moving average of squared gradient values
state['exp_avg_var'] = torch.zeros_like(p.data,
memory_format=torch.preserve_format) if version_higher else torch.zeros_like(p.data)
# Maintains max of all exp. moving avg. of sq. grad. values
state['max_exp_avg_var'] = torch.zeros_like(p.data,
memory_format=torch.preserve_format) if version_higher else torch.zeros_like(p.data)

# get current state variable
exp_avg, exp_avg_var = state['exp_avg'], state['exp_avg_var']

state['step'] += 1
bias_correction1 = 1 - beta1 ** state['step']
bias_correction2 = 1 - beta2 ** state['step']

# perform weight decay, check if decoupled weight decay
if self.weight_decouple:
if not self.fixed_decay:
p.data.mul_(1.0 - group['lr'] * group['weight_decay'])
else:
p.data.mul_(1.0 - group['weight_decay'])
else:
if group['weight_decay'] != 0:

# Update first and second moment running average

max_exp_avg_var = state['max_exp_avg_var']
# Maintains the maximum of all 2nd moment running avg. till now
torch.max(max_exp_avg_var, exp_avg_var, out=max_exp_avg_var)

# Use the max. for normalizing running avg. of gradient
else:

if not self.rectify:
# Default update
step_size = group['lr'] / bias_correction1

else:# Rectified update
# calculate rho_t
state['rho_t'] = state['rho_inf'] - 2 * state['step'] * beta2 ** state['step'] / (
1.0 - beta2 ** state['step'])

if state['rho_t'] > 4: # perform Adam style update if variance is small
rho_inf, rho_t = state['rho_inf'], state['rho_t']
rt = (rho_t - 4.0) * (rho_t - 2.0) * rho_inf / (rho_inf - 4.0) / (rho_inf - 2.0) / rho_t
rt = math.sqrt(rt)

step_size = rt * group['lr'] / bias_correction1

else: # perform SGD style update

return loss



0
0 收藏

0 评论
0 收藏
0