# Copyright 2016-2023 Cerebras Systems
# SPDX-License-Identifier: BSD-3-Clause
"""contains the Cerebras Adadelta implementation"""
from typing import Callable
import torch
import cerebras.pytorch as cstorch
from .optimizer import Optimizer
[docs]class Adadelta(Optimizer):
"""
Adadelta optimizer implemented to perform the required
pre-initialization of the optimizer state.
"""
def __init__(
self,
params,
lr=1.0,
rho=0.9,
eps=1e-6,
weight_decay=0,
maximize: bool = False,
):
if lr < 0.0:
raise ValueError("Invalid learning rate: {}".format(lr))
if not 0.0 <= rho <= 1.0:
raise ValueError("Invalid rho value: {}".format(rho))
if eps < 0.0:
raise ValueError("Invalid epsilon value: {}".format(eps))
if weight_decay < 0.0:
raise ValueError(
"Invalid weight_decay value: {}".format(weight_decay)
)
defaults = dict(
lr=lr,
rho=rho,
eps=eps,
weight_decay=weight_decay,
maximize=maximize,
)
super().__init__(params, defaults)
[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]["square_avg"] = cstorch.zeros_like(p)
self.state[p]["acc_delta"] = cstorch.zeros_like(p)
@torch.no_grad()
def step(self, closure: Callable = None):
"""Performs a single optimization step.
Args:
closure : 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"]
rho = group['rho']
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(
"Adadelta does not support sparse gradients."
)
state = self.state[p]
square_avg = state["square_avg"]
acc_delta = state["acc_delta"]
grad = grad if not maximize else -grad
grad = grad + p * weight_decay
square_avg.mul_(rho).addcmul_(grad, grad, value=1 - rho)
std = square_avg.add(eps).sqrt_()
delta = acc_delta.add(eps).sqrt_().div_(std).mul_(grad)
acc_delta.mul_(rho).addcmul_(delta, delta, value=1 - rho)
p.add_(-lr * delta)
return loss