bitsandbytes-rocm/bitsandbytes/optim/adagrad.py
Tom Aarsen 0b078403ee Simplify statements into equivalent, modern variants
via pyupgrade --py37-plus. The changes e.g. are subclassing from object, calling super() with super(ThisClass, self), or old-style syntax formatting.
2022-10-27 13:14:13 +02:00

133 lines
3.6 KiB
Python

# 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.
from bitsandbytes.optim.optimizer import Optimizer1State
class Adagrad(Optimizer1State):
def __init__(
self,
params,
lr=1e-2,
lr_decay=0,
weight_decay=0,
initial_accumulator_value=0,
eps=1e-10,
optim_bits=32,
args=None,
min_8bit_size=4096,
percentile_clipping=100,
block_wise=True,
):
if not 0.0 <= lr:
raise ValueError(f"Invalid learning rate: {lr}")
if not 0.0 <= weight_decay:
raise ValueError(
f"Invalid weight_decay value: {weight_decay}"
)
if not 0.0 <= eps:
raise ValueError(f"Invalid epsilon value: {eps}")
if initial_accumulator_value != 0.0:
raise ValueError("Initial accumulator value != 0.0 not supported!")
if lr_decay != 0.0:
raise ValueError("Lr Decay != 0.0 not supported!")
super().__init__(
"adagrad",
params,
lr,
(0.0, 0.0),
eps,
weight_decay,
optim_bits,
args,
min_8bit_size,
percentile_clipping,
block_wise,
)
class Adagrad8bit(Optimizer1State):
def __init__(
self,
params,
lr=1e-2,
lr_decay=0,
weight_decay=0,
initial_accumulator_value=0,
eps=1e-10,
optim_bits=8,
args=None,
min_8bit_size=4096,
percentile_clipping=100,
block_wise=True,
):
if not 0.0 <= lr:
raise ValueError(f"Invalid learning rate: {lr}")
if not 0.0 <= weight_decay:
raise ValueError(
f"Invalid weight_decay value: {weight_decay}"
)
if not 0.0 <= eps:
raise ValueError(f"Invalid epsilon value: {eps}")
if initial_accumulator_value != 0.0:
raise ValueError("Initial accumulator value != 0.0 not supported!")
if lr_decay != 0.0:
raise ValueError("Lr Decay != 0.0 not supported!")
assert block_wise
super().__init__(
"adagrad",
params,
lr,
(0.0, 0.0),
eps,
weight_decay,
8,
args,
min_8bit_size,
percentile_clipping,
block_wise,
)
class Adagrad32bit(Optimizer1State):
def __init__(
self,
params,
lr=1e-2,
lr_decay=0,
weight_decay=0,
initial_accumulator_value=0,
eps=1e-10,
optim_bits=32,
args=None,
min_8bit_size=4096,
percentile_clipping=100,
block_wise=True,
):
if not 0.0 <= lr:
raise ValueError(f"Invalid learning rate: {lr}")
if not 0.0 <= weight_decay:
raise ValueError(
f"Invalid weight_decay value: {weight_decay}"
)
if not 0.0 <= eps:
raise ValueError(f"Invalid epsilon value: {eps}")
if initial_accumulator_value != 0.0:
raise ValueError("Initial accumulator value != 0.0 not supported!")
if lr_decay != 0.0:
raise ValueError("Lr Decay != 0.0 not supported!")
super().__init__(
"adagrad",
params,
lr,
(0.0, 0.0),
eps,
weight_decay,
32,
args,
min_8bit_size,
percentile_clipping,
block_wise,
)