import math
import torch
[docs]
class AdamW(torch.optim.Optimizer):
r"""Implements AdamW algorithm. This code is adapated from `PyTorch codebase <https://github.com/pytorch/pytorch/blob/v1.2.0/torch/optim/adamw.py>`__.
The original Adam algorithm was proposed in `Adam: A Method for Stochastic Optimization`_.
The AdamW variant was proposed in `Decoupled Weight Decay Regularization`_.
Arguments:
params (iterable): iterable of parameters to optimize
lr (float): 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 coefficient (default: ``1e-2``)
epoch_decay (float, optional): epoch decay (epoch-wise l2 penalty) (default: ``0.0``)
amsgrad (boolean, optional): whether to use the AMSGrad variant of this
algorithm from the paper `On the Convergence of Adam and Beyond`_
(default: False)
device (torch.device, optional): the device used for optimization, e.g., 'cpu' or 'cuda' (default: ``None``).
.. _Adam\: A Method for Stochastic Optimization:
https://arxiv.org/abs/1412.6980
.. _Decoupled Weight Decay Regularization:
https://arxiv.org/abs/1711.05101
.. _On the Convergence of Adam and Beyond:
https://openreview.net/forum?id=ryQu7f-RZ
"""
def __init__(self,
params,
lr=1e-3,
betas=(0.9, 0.999),
eps=1e-8,
clip_value=10.0,
epoch_decay=0,
weight_decay=1e-2,
amsgrad=False,
verbose=False,
device=None,
**kwargs):
if not device:
self.device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
else:
self.device = device
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]))
if not 0.0 <= weight_decay:
raise ValueError("Invalid weight_decay value: {}".format(weight_decay))
self.params = list(params)
self.lr = lr
self.model_ref = self.__init_model_ref__(self.params) if epoch_decay > 0 else None
self.model_acc = self.__init_model_acc__(self.params) if epoch_decay > 0 else None
self.T = 0 # for epoch_decay
self.steps = 0 # total optim steps
self.verbose = verbose # print updates for lr/regularizer
defaults = dict(lr=lr, betas=betas, eps=eps,
weight_decay=weight_decay, epoch_decay=epoch_decay, amsgrad=amsgrad,
clip_value=clip_value, model_ref=self.model_ref, model_acc=self.model_acc)
super(AdamW, self).__init__(self.params, defaults)
def __setstate__(self, state):
super(AdamW, self).__setstate__(state)
for group in self.param_groups:
group.setdefault('amsgrad', False)
def __init_model_ref__(self, params):
model_ref = []
if not isinstance(params, list):
params = list(params)
for var in params:
if var is not None:
model_ref.append(torch.empty(var.shape).normal_(mean=0, std=0.01).to(self.device))
return model_ref
def __init_model_acc__(self, params):
model_acc = []
if not isinstance(params, list):
params = list(params)
for var in params:
if var is not None:
model_acc.append(torch.zeros(var.shape, dtype=torch.float32, device=self.device, requires_grad=False).to(self.device))
return model_acc
[docs]
@torch.no_grad()
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:
with torch.enable_grad():
loss = closure()
for group in self.param_groups:
self.lr = group['lr']
model_ref = group['model_ref']
model_acc = group['model_acc']
epoch_decay = group['epoch_decay']
clip_value = group['clip_value']
for i, p in enumerate(group['params']):
if p.grad is None:
continue
# Perform stepweight decay
p.mul_(1 - self.lr * group['weight_decay'])
# Perform optimization step
if epoch_decay > 0:
grad = torch.clamp(p.grad.data , -clip_value, clip_value) + epoch_decay*(p.data - model_ref[i].data) #p.grad
else:
grad = torch.clamp(p.grad.data , -clip_value, clip_value)
if grad.is_sparse:
raise RuntimeError('AdamW does not support sparse gradients')
amsgrad = group['amsgrad']
state = self.state[p]
# State initialization
if len(state) == 0:
state['step'] = 0
# Exponential moving average of gradient values
state['exp_avg'] = torch.zeros_like(p, memory_format=torch.preserve_format)
# Exponential moving average of squared gradient values
state['exp_avg_sq'] = torch.zeros_like(p, memory_format=torch.preserve_format)
if amsgrad:
# Maintains max of all exp. moving avg. of sq. grad. values
state['max_exp_avg_sq'] = torch.zeros_like(p, memory_format=torch.preserve_format)
exp_avg, exp_avg_sq = state['exp_avg'], state['exp_avg_sq']
if amsgrad:
max_exp_avg_sq = state['max_exp_avg_sq']
beta1, beta2 = group['betas']
state['step'] += 1
bias_correction1 = 1 - beta1 ** state['step']
bias_correction2 = 1 - beta2 ** state['step']
# Decay the first and second moment running average coefficient
exp_avg.mul_(beta1).add_(grad, alpha=1 - beta1)
exp_avg_sq.mul_(beta2).addcmul_(grad, grad, value=1 - beta2)
if amsgrad:
# Maintains the maximum of all 2nd moment running avg. till now
torch.maximum(max_exp_avg_sq, exp_avg_sq, out=max_exp_avg_sq)
# Use the max. for normalizing running avg. of gradient
denom = (max_exp_avg_sq.sqrt() / math.sqrt(bias_correction2)).add_(group['eps'])
else:
denom = (exp_avg_sq.sqrt() / math.sqrt(bias_correction2)).add_(group['eps'])
step_size = self.lr / bias_correction1
p.addcdiv_(exp_avg, denom, value=-step_size)
if epoch_decay > 0:
model_acc[i].data = model_acc[i].data + p.data
self.T += 1
self.steps += 1
return loss
[docs]
def update_lr(self, decay_factor=None):
r"""Updates learning rate given a decay factor."""
if decay_factor != None:
self.param_groups[0]['lr'] = self.param_groups[0]['lr']/decay_factor
if self.verbose:
print ('Reducing learning rate to %.5f !'%(self.param_groups[0]['lr']))
[docs]
def update_regularizer(self, decay_factor=None):
r"""Updates learning rate given a decay factor and resets epoch-decay regularizer."""
if decay_factor != None:
self.param_groups[0]['lr'] = self.param_groups[0]['lr']/decay_factor
if self.verbose:
print ('Reducing learning rate to %.5f @ T=%s!'%(self.param_groups[0]['lr'], self.steps))
if self.verbose:
print ('Updating regularizer @ T=%s!'%(self.steps))
for i, param in enumerate(self.model_ref):
self.model_ref[i].data = self.model_acc[i].data/self.T
for i, param in enumerate(self.model_acc):
self.model_acc[i].data = torch.zeros(param.shape, dtype=torch.float32, device=self.device, requires_grad=False).to(self.device)
self.T = 0