#
# Cerebras implementation of ASGD optimizer. Adapted from the `torch.optim.ASGD` implementation.
#
# Copyright 2016-2023 Cerebras Systems
# SPDX-License-Identifier: BSD-3-Clause
#
import torch
import cerebras.pytorch as cstorch
from .optimizer import Optimizer
[docs]class ASGD(Optimizer):
r"""ASGD optimizer implemented to conform to execution within the constraints
of the Cerebras WSE, including pre-initializing optimizer state.
For more details, see https://dl.acm.org/citation.cfm?id=131098
"""
def __init__(
self,
params,
lr=1e-2,
lambd=1e-4,
alpha=0.75,
t0=1e6,
weight_decay=0,
maximize: bool = False,
):
if lr < 0.0:
raise ValueError(f"Invalid learning rate: {lr}")
if weight_decay < 0.0:
raise ValueError(f"Invalid weight_decay value: {weight_decay}")
defaults = dict(
lr=lr,
lambd=lambd,
alpha=alpha,
t0=t0,
weight_decay=weight_decay,
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]["eta"] = torch.tensor(group["lr"]).to(p.device)
self.state[p]["mu"] = torch.tensor(1.0).to(p.device)
self.state[p]["ax"] = cstorch.zeros_like(p)
@torch.no_grad()
def step(self, closure=None):
r"""
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:
with torch.enable_grad():
loss = closure()
for group in self.param_groups:
lambd = group["lambd"]
lr = group["lr"]
t0 = group["t0"]
for p in group["params"]:
if p.grad is not None:
if p.grad.is_sparse:
raise RuntimeError(
"ASGD does not support sparse gradients"
)
alpha = group["alpha"]
if not isinstance(alpha, torch.Tensor):
alpha = torch.tensor(alpha)
alpha = alpha.to(p.device)
state = self.state[p]
grad = p.grad
grad = grad if not group["maximize"] else -grad
mu = state["mu"]
ax = state["ax"]
eta = state["eta"]
step = self.increment_global_step(p)
grad = grad + p * group["weight_decay"]
# decay term
p.mul_(1 - lambd * eta)
# update parameter
p.add_(grad * eta.neg())
# averaging
new_ax = torch.where(mu == 1, p, ax.add(p.sub(ax).mul(mu)))
ax.copy_(new_ax)
new_eta = lr / torch.pow(1 + lambd * lr * step, alpha)
eta.copy_(new_eta)
new_mu = 1 / torch.maximum(
torch.ones(size=[], dtype=mu.dtype),
step - t0,
)
mu.copy_(new_mu)
return loss