# Copyright (c) Facebook, Inc. and its affiliates.
#
# This source code is licensed under the MIT license found in the
# LICENSE file in the root directory of this source tree.
import math
import types
import torch
import torch.optim
from . import FairseqOptimizer, register_optimizer
[docs]@register_optimizer('adam')
class FairseqAdam(FairseqOptimizer):
def __init__(self, args, params):
super().__init__(args, params)
if torch.cuda.is_available():
try:
from apex.optimizers import FusedAdam as _FusedAdam # noqa
self._optimizer = FusedAdam(params, **self.optimizer_config)
except ImportError:
self._optimizer = Adam(params, **self.optimizer_config)
else:
self._optimizer = Adam(params, **self.optimizer_config)
[docs] @staticmethod
def add_args(parser):
"""Add optimizer-specific arguments to the parser."""
# fmt: off
parser.add_argument('--adam-betas', default='(0.9, 0.999)', metavar='B',
help='betas for Adam optimizer')
parser.add_argument('--adam-eps', type=float, default=1e-8, metavar='D',
help='epsilon for Adam optimizer')
parser.add_argument('--weight-decay', '--wd', default=0.0, type=float, metavar='WD',
help='weight decay')
# fmt: on
@property
def optimizer_config(self):
"""
Return a kwarg dictionary that will be used to override optimizer
args stored in checkpoints. This allows us to load a checkpoint and
resume training using a different set of optimizer args, e.g., with a
different learning rate.
"""
return {
'lr': self.args.lr[0],
'betas': eval(self.args.adam_betas),
'eps': self.args.adam_eps,
'weight_decay': self.args.weight_decay,
}
class Adam(torch.optim.Optimizer):
"""Implements Adam algorithm.
This implementation is modified from torch.optim.Adam based on:
`Fixed Weight Decay Regularization in Adam`
(see https://arxiv.org/abs/1711.05101)
It has been proposed in `Adam: A Method for Stochastic Optimization`_.
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)
amsgrad (boolean, optional): whether to use the AMSGrad variant of this
algorithm from the paper `On the Convergence of Adam and Beyond`_
.. _Adam\: A Method for Stochastic Optimization:
https://arxiv.org/abs/1412.6980
.. _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,
weight_decay=0, amsgrad=False):
defaults = dict(lr=lr, betas=betas, eps=eps,
weight_decay=weight_decay, amsgrad=amsgrad)
super(Adam, self).__init__(params, defaults)
@property
def supports_memory_efficient_fp16(self):
return True
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']:
if p.grad is None:
continue
grad = p.grad.data.float()
if grad.is_sparse:
raise RuntimeError('Adam does not support sparse gradients, please consider SparseAdam instead')
amsgrad = group['amsgrad']
p_data_fp32 = p.data.float()
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_data_fp32)
# Exponential moving average of squared gradient values
state['exp_avg_sq'] = torch.zeros_like(p_data_fp32)
if amsgrad:
# Maintains max of all exp. moving avg. of sq. grad. values
state['max_exp_avg_sq'] = torch.zeros_like(p_data_fp32)
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)
if amsgrad:
state['max_exp_avg_sq'] = state['max_exp_avg_sq'].type_as(p_data_fp32)
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
# Decay the first and second moment running average coefficient
exp_avg.mul_(beta1).add_(1 - beta1, grad)
exp_avg_sq.mul_(beta2).addcmul_(1 - beta2, grad, grad)
if amsgrad:
# Maintains the maximum of all 2nd moment running avg. till now
torch.max(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().add_(group['eps'])
else:
denom = exp_avg_sq.sqrt().add_(group['eps'])
bias_correction1 = 1 - beta1 ** state['step']
bias_correction2 = 1 - beta2 ** state['step']
step_size = group['lr'] * math.sqrt(bias_correction2) / bias_correction1
if group['weight_decay'] != 0:
p_data_fp32.add_(-group['weight_decay'] * group['lr'], p_data_fp32)
p_data_fp32.addcdiv_(-step_size, exp_avg, denom)
p.data.copy_(p_data_fp32)
return loss
class FusedAdam(torch.optim.Optimizer):
"""
Implements Adam algorithm. Currently GPU-only. Requires Apex to be installed via
``python setup.py install --cuda_ext --cpp_ext``.
It has been proposed in `Adam: A Method for Stochastic Optimization`_.
Compared to the original version in Apex, the fairseq version casts grads
and params to FP32 internally to support ``--memory-efficient-fp16``.
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)
amsgrad (boolean, optional): whether to use the AMSGrad variant of this
algorithm from the paper `On the Convergence of Adam and Beyond`_
(default: False) NOT SUPPORTED in FusedAdam!
eps_inside_sqrt (boolean, optional): in the 'update parameters' step,
adds eps to the bias-corrected second moment estimate before
evaluating square root instead of adding it to the square root of
second moment estimate as in the original paper. (default: False)
.. _Adam\: A Method for Stochastic Optimization:
https://arxiv.org/abs/1412.6980
.. _On the Convergence of Adam and Beyond:
https://openreview.net/forum?id=ryQu7f-RZ
"""
def __init__(self, params,
lr=1e-3, bias_correction = True,
betas=(0.9, 0.999), eps=1e-8, eps_inside_sqrt=False,
weight_decay=0., max_grad_norm=0., amsgrad=False):
global fused_adam_cuda
import importlib
fused_adam_cuda = importlib.import_module("fused_adam_cuda")
if amsgrad:
raise RuntimeError('FusedAdam does not support the AMSGrad variant.')
defaults = dict(lr=lr, bias_correction=bias_correction,
betas=betas, eps=eps, weight_decay=weight_decay,
max_grad_norm=max_grad_norm)
super(FusedAdam, self).__init__(params, defaults)
self.eps_mode = 0 if eps_inside_sqrt else 1
@property
def supports_memory_efficient_fp16(self):
return True
def step(self, closure=None, grads=None, scale=1., grad_norms=None):
"""Performs a single optimization step.
Arguments:
closure (callable, optional): A closure that reevaluates the model
and returns the loss.
grads (list of tensors, optional): weight gradient to use for the
optimizer update. If gradients have type torch.half, parameters
are expected to be in type torch.float. (default: None)
output params (list of tensors, optional): A reduced precision copy
of the updated weights written out in addition to the regular
updated weights. Have to be of same type as gradients. (default: None)
scale (float, optional): factor to divide gradient tensor values
by before applying to weights. (default: 1)
"""
loss = None
if closure is not None:
loss = closure()
if grads is None:
grads_group = [None]*len(self.param_groups)
# backward compatibility
# assuming a list/generator of parameter means single group
elif isinstance(grads, types.GeneratorType):
grads_group = [grads]
elif type(grads[0])!=list:
grads_group = [grads]
else:
grads_group = grads
if grad_norms is None:
grad_norms = [None]*len(self.param_groups)
for group, grads_this_group, grad_norm in zip(self.param_groups, grads_group, grad_norms):
if grads_this_group is None:
grads_this_group = [None]*len(group['params'])
# compute combined scale factor for this group
combined_scale = scale
if group['max_grad_norm'] > 0:
# norm is in fact norm*scale
clip = ((grad_norm / scale) + 1e-6) / group['max_grad_norm']
if clip > 1:
combined_scale = clip * scale
bias_correction = 1 if group['bias_correction'] else 0
for p, grad in zip(group['params'], grads_this_group):
#note: p.grad should not ever be set for correct operation of mixed precision optimizer that sometimes sends None gradients
if p.grad is None and grad is None:
continue
if grad is None:
grad = p.grad.data
if grad.is_sparse:
raise RuntimeError('FusedAdam does not support sparse gradients, please consider SparseAdam instead')
p_data_fp32 = p.data.float()
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_data_fp32)
# Exponential moving average of squared gradient values
state['exp_avg_sq'] = torch.zeros_like(p_data_fp32)
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)
exp_avg = state['exp_avg']
exp_avg_sq = state['exp_avg_sq']
beta1, beta2 = group['betas']
state['step'] += 1
out_p = p.data
fused_adam_cuda.adam(p_data_fp32,
out_p,
exp_avg,
exp_avg_sq,
grad,
group['lr'],
beta1,
beta2,
group['eps'],
combined_scale,
state['step'],
self.eps_mode,
bias_correction,
group['weight_decay'])
return loss