import torch.optim as optim
from torch.nn.utils import clip_grad_norm_
class Optim(object):
"""
https://github.com/OpenNMT/OpenNMT-py/blob/master/onmt/
utils/optimizers.py
Controller class for optimization. Mostly a thin
wrapper for `optim`, but also useful for implementing
rate scheduling beyond what is currently available.
Also implements necessary methods for training RNNs such
as grad manipulations.
Args:
method (:obj:`str`): one of [sgd, adagrad, adadelta, adam]
lr (float): learning rate
lr_decay (float, optional): learning rate decay multiplier
start_decay_at (int, optional): epoch to start learning rate decay
beta1, beta2 (float, optional): parameters for adam
adagrad_accum (float, optional): initialization parameter for adagrad
decay_method (str, option): custom decay options
warmup_steps (int, option): parameter for `noam` decay
model_size (int, option): parameter for `noam` decay
"""
# We use the default parameters for Adam that are suggested by
# the original paper https://arxiv.org/pdf/1412.6980.pdf
# These values are also used by other established implementations,
# e.g. https://www.tensorflow.org/api_docs/python/tf/train/AdamOptimizer
# https://keras.io/optimizers/
# Recently there are slightly different values used in the paper
# "Attention is all you need"
# https://arxiv.org/pdf/1706.03762.pdf, particularly the value beta2=0.98
# was used there however, beta2=0.999 is still arguably the more
# established value, so we use that here as well
def __init__(self, method, learning_rate, max_grad_norm,
lr_decay=1, start_decay_steps=None, decay_steps=None,
beta1=0.9, beta2=0.999, adam_eps=1e-9,
adagrad_accum=0.0,
decay_method=None,
warmup_steps=4000,
model_size=None):
self.last_ppl = None
self.learning_rate = learning_rate
self.original_lr = learning_rate
self.max_grad_norm = max_grad_norm
self.method = method
self.lr_decay = lr_decay
self.start_decay_steps = start_decay_steps
self.decay_steps = decay_steps
self.start_decay = False
self._step = 0
self.betas = [beta1, beta2]
self.adam_eps = adam_eps
self.adagrad_accum = adagrad_accum
self.decay_method = decay_method
self.warmup_steps = warmup_steps
self.model_size = model_size
def set_parameters(self, params):
""" ? """
self.params = []
self.sparse_params = []
for k, p in params:
if p.requires_grad:
if self.method != 'sparseadam' or "embed" not in k:
self.params.append(p)
else:
self.sparse_params.append(p)
if self.method == 'sgd':
self.optimizer = optim.SGD(self.params, lr=self.learning_rate)
elif self.method == 'adagrad':
self.optimizer = optim.Adagrad(self.params, lr=self.learning_rate)
for group in self.optimizer.param_groups:
for p in group['params']:
self.optimizer.state[p]['sum'] = self.optimizer\
.state[p]['sum'].fill_(self.adagrad_accum)
elif self.method == 'adadelta':
self.optimizer = optim.Adadelta(self.params, lr=self.learning_rate)
elif self.method == 'adam':
self.optimizer = optim.Adam(self.params, lr=self.learning_rate,
betas=self.betas, eps=self.adam_eps)
else:
raise RuntimeError("Invalid optim method: " + self.method)
def _set_rate(self, learning_rate):
self.learning_rate = learning_rate
self.optimizer.param_groups[0]['lr'] = self.learning_rate
def step(self):
"""Update the model parameters based on current gradients.
Optionally, will employ gradient modification or update learning
rate.
"""
self._step += 1
# Decay method used in tensor2tensor.
if self.decay_method == "noam":
self._set_rate(
self.original_lr *
(self.model_size ** (-0.5) *
min(self._step ** (-0.5),
self._step * self.warmup_steps**(-1.5))))
# Decay based on start_decay_steps every decay_steps
else:
if ((self.start_decay_steps is not None) and (
self._step >= self.start_decay_steps)):
self.start_decay = True
if self.start_decay:
if ((self._step - self.start_decay_steps)
% self.decay_steps == 0):
self.learning_rate = self.learning_rate * self.lr_decay
if self.method != 'sparseadam':
self.optimizer.param_groups[0]['lr'] = self.learning_rate
if self.max_grad_norm:
clip_grad_norm_(self.params, self.max_grad_norm)
self.optimizer.step()
