# Copyright 2016-2023 Cerebras Systems
# SPDX-License-Identifier: BSD-3-Clause
"""contains the Cerebras Adagrad implementation."""
import torch
import cerebras.pytorch as cstorch
from .optimizer import Optimizer, ParamsT
[docs]class Adagrad(Optimizer):
r"""Adagrad optimizer implemented to conform to execution within the
constraints of the Cerebras WSE.
Args:
params: iterable of parameters to optimize or dicts defining parameter groups
lr: learning rate
lr_decay: learning rate decay
weight_decay: weight decay (L2 penalty)
eps: term added to the denominator to improve numerical stability
maximize: maximize the params based on the objective instead of minimizing
Adaptive Subgradient Methods for Online Learning and Stochastic
Optimization: http://jmlr.org/papers/v12/duchi11a.html
"""
def __init__(
self,
params: ParamsT,
lr: float = 1e-2,
lr_decay: float = 0,
weight_decay: float = 0,
initial_accumulator_value: float = 0,
eps: float = 1e-6,
maximize: bool = False,
):
if lr < 0.0:
raise ValueError("Invalid learning rate: {}".format(lr))
if lr_decay < 0.0:
raise ValueError("Invalid lr_decay value: {}".format(lr_decay))
if weight_decay < 0.0:
raise ValueError(
"Invalid weight_decay value: {}".format(weight_decay)
)
if initial_accumulator_value < 0.0:
raise ValueError(
"Invalid initial_accumulator_value value: {}".format(
initial_accumulator_value
)
)
if eps < 0.0:
raise ValueError("Invalid epsilon value: {}".format(eps))
defaults = dict(
lr=lr,
lr_decay=lr_decay,
eps=eps,
weight_decay=weight_decay,
initial_accumulator_value=initial_accumulator_value,
maximize=maximize,
)
super().__init__(params, defaults, enable_global_step=True)
[docs] def preinitialize(self):
"""
Allocates tensors for the optimizer state to allow direct compilation
of the model before the first step.
"""
for group in self.param_groups:
for p in group['params']:
self.state[p]["sum"] = cstorch.full_like(
p,
group["initial_accumulator_value"],
)
@torch.no_grad()
def step(self, closure=None):
"""Performs a single optimization step.
Args:
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:
lr = group["lr"]
weight_decay = group["weight_decay"]
lr_decay = group["lr_decay"]
eps = group["eps"]
maximize = group["maximize"]
for p in group["params"]:
if p.grad is None:
continue
grad = p.grad
if grad.is_sparse:
raise RuntimeError(
"Adagrad does not support sparse gradients."
)
state = self.state[p]
state_sum = state["sum"]
global_step = self.increment_global_step(p)
grad = grad if not maximize else -grad
grad = grad + p * weight_decay
state_sum.addcmul_(grad, grad, value=1.0)
std = state_sum.sqrt().add_(eps)
# BEGIN_CEREBRAS_ONLY
# The following two lines implements clr, in two steps:
# clr = lr / (1.0 + (global_step - 1.0) * lr_decay)
# This workaround avoids LR constant folding?
# END_CEREBRAS_ONLY
grad.div_(1.0 + (global_step - 1.0) * lr_decay)
p.addcdiv_(-lr * grad, std)
return loss