paddlepaddle--paddle/python/paddle/optimizer/lamb.py

# Copyright (c) 2022 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

from __future__ import annotations

from typing import TYPE_CHECKING

from paddle import _C_ops, pir
from paddle.base.executor import global_scope

from ..base import core, framework
from ..base.framework import Variable
from .optimizer import Optimizer

if TYPE_CHECKING:
    from collections.abc import Callable, Sequence

    from typing_extensions import NotRequired

    from paddle import Tensor
    from paddle.nn.clip import GradientClipBase

    from .optimizer import _ParameterConfig

    class _LambParameterConfig(_ParameterConfig):
        beta1: NotRequired[float | Tensor]
        beta2: NotRequired[float | Tensor]
        epsilon: NotRequired[float | Tensor]
        lamb_weight_decay: NotRequired[float]
        exclude_from_weight_decay_fn: NotRequired[
            Callable[[Tensor], bool] | None
        ]


__all__ = []


class Lamb(Optimizer):
    r"""
    LAMB (Layer-wise Adaptive Moments optimizer for Batching training) Optimizer.

    LAMB Optimizer is designed to scale up the batch size of training without losing
    accuracy, which supports adaptive element-wise updating and accurate layer-wise
    correction. For more information, please refer to `Large Batch Optimization for
    Deep Learning: Training BERT in 76 minutes <https://arxiv.org/abs/1904.00962>`_ .

    The updating of parameters follows:

    ..  math::

        m_t &= \beta_1 m_{t - 1}+ (1 - \beta_1)g_t

        v_t &= \beta_2 v_{t - 1}  + (1 - \beta_2)g_t^2

        m_t &= \frac{m_t}{\beta_1^t}

        v_t &= \frac{v_t}{\beta_2^t}

        r_t &= \frac{m_t}{\sqrt{v_t}+\epsilon}

        w_t &= w_{t-1} -\eta_t \frac{\left \| w_{t-1}\right \|}{\left \| r_t + \lambda w_{t-1}\right \|} (r_t + \lambda w_{t-1})


    where :math:`m` is the 1st moment, and :math:`v` the 2nd moment, :math:`\\eta` the
    learning rate, :math:`\\lambda` the LAMB weight decay rate.

    Args:
        learning_rate (float|Tensor, optional): the learning rate used to update parameters. \
            Can be a float value or a Variable with data type float32. Default 0.001.
        lamb_weight_decay (float, optional): The LAMB weight decay rate. Default 0.01. Remind that weight_decay should be None.
        beta1 (float|Tensor, optional): The exponential decay rate for the 1st moment estimates.
            Default 0.9.
        beta2 (float|Tensor, optional): The exponential decay rate for the 2nd moment estimates.
            Default 0.999.
        epsilon (float|Tensor, optional): A small float value for numerical stability. Default 1e-6.
        parameters (list|tuple|None, optional):  Iterable of ``Variable`` names to update to minimize ``loss``. \
            This parameter is required in dygraph mode. And you can specify different options for \
            different parameter groups such as the learning rate, weight decay, etc, \
            then the parameters are list of dict. Note that the learning_rate in parameter groups \
            represents the scale of base learning_rate. \
            The default value is None in static graph mode, at this time all parameters will be updated.
        grad_clip (GradientClipBase, optional): Gradient clipping strategy, it's an instance of
            some derived class of ``GradientClipBase`` . There are three clipping strategies
            ( :ref:`api_paddle_base_clip_ClipGradByGlobalNorm` , :ref:`api_paddle_base_clip_ClipGradByNorm` ,
            :ref:`api_paddle_base_clip_ClipGradByValue` ). If you want better convergence, it is recommended
            to use :ref:`api_paddle_base_clip_ClipGradByGlobalNorm` . Default None, meaning there is no gradient clipping.
        exclude_from_weight_decay_fn (Callable|None, optional): whether to skip weight decay for a parameter when this function returns True while take the parameter as input.
        multi_precision (bool, optional) - Whether to use it during weight updates multi-precision, Default False。
        always_adapt (bool, optional): whether to use Layer-wise LR adaptation. By default, skip adaptation on parameters that are
            excluded from weight decay, unless always_adapt == True, then always enable LR adaptation.
        name(str|None, optional): For detailed information, please refer to
            :ref:`api_guide_Name` . Usually name is no need to set and None by default.
    Examples:
        .. code-block:: pycon

            >>> import paddle

            >>> inp = paddle.uniform(shape=[10, 10], dtype='float32', min=-0.1, max=0.1)
            >>> linear = paddle.nn.Linear(10, 10)
            >>> out = linear(inp)
            >>> loss = paddle.mean(out)
            >>> beta1 = paddle.to_tensor([0.9], dtype="float32")
            >>> beta2 = paddle.to_tensor([0.85], dtype="float32")
            >>> lamb = paddle.optimizer.Lamb(
            ...     learning_rate=0.002,
            ...     beta1=beta1,
            ...     beta2=beta2,
            ...     parameters=linear.parameters(),
            ...     lamb_weight_decay=0.01
            ... )
            >>> back = out.backward()
            >>> lamb.step()
            >>> lamb.clear_grad()

    """

    _moment1_acc_str = "moment1"
    _moment2_acc_str = "moment2"
    _beta1_pow_acc_str = "beta1_pow_acc"
    _beta2_pow_acc_str = "beta2_pow_acc"

    def __init__(
        self,
        learning_rate: float | Tensor = 0.001,
        lamb_weight_decay: float = 0.01,
        beta1: float | Tensor = 0.9,
        beta2: float | Tensor = 0.999,
        epsilon: float | Tensor = 1e-6,
        parameters: (
            Sequence[Tensor] | Sequence[_LambParameterConfig] | None
        ) = None,
        grad_clip: GradientClipBase | None = None,
        exclude_from_weight_decay_fn: Callable[[Tensor], bool] | None = None,
        multi_precision: bool = False,
        always_adapt: bool = False,
        name: str | None = None,
    ) -> None:
        assert learning_rate is not None
        assert beta1 is not None
        assert beta2 is not None
        assert epsilon is not None
        super().__init__(
            learning_rate=learning_rate,
            parameters=parameters,
            weight_decay=None,
            grad_clip=grad_clip,
            name=name,
        )
        self.type = "lamb"
        self._beta1 = beta1
        self._beta2 = beta2
        self._epsilon = epsilon
        self._lamb_weight_decay = lamb_weight_decay
        self._exclude_from_weight_decay_fn = exclude_from_weight_decay_fn
        self._default_dict = {
            'beta1': beta1,
            'beta2': beta2,
            'epsilon': epsilon,
            'lamb_weight_decay': lamb_weight_decay,
            'exclude_from_weight_decay_fn': exclude_from_weight_decay_fn,
        }
        self._master_weights = {}
        self._used_master_weights = {}
        # TODO(zengjinle): expose API as soon as possible
        self._multi_precision = multi_precision
        self.always_adapt = always_adapt

    def _get_parameter(self, name, scope=None):
        if scope is None:
            scope = global_scope()

        p_t = scope.find_var(name).get_tensor()

        master_name = self._used_master_weights.get(name)
        if master_name is not None:
            master_p_t = scope.find_var(master_name).get_tensor()
            assert master_p_t._dtype() != p_t._dtype()
            assert master_p_t.shape() == p_t.shape()
        else:
            master_p_t = None
        return p_t, master_p_t

    def _create_accumulators(self, block, parameters):
        if not isinstance(block, (framework.Block, pir.Block)):
            raise TypeError("block is not instance of Block.")
        if isinstance(parameters, dict):
            parameters = self._update_param_group(parameters)

        # Create accumulator tensors for first and second moments
        for p in parameters:
            if p.name in self._already_create_accumulator:
                continue
            if self._multi_precision and self._is_dtype_fp16_or_bf16(p.dtype):
                master_p = self._create_master_weight(p)
                self._add_moments_pows(master_p)
                self._already_create_accumulator.add(p.name)
            else:
                self._add_moments_pows(p)
                self._already_create_accumulator.add(p.name)

    def _add_moments_pows(self, p):
        acc_dtype = p.dtype
        if self._is_dtype_fp16_or_bf16(acc_dtype):
            acc_dtype = core.VarDesc.VarType.FP32

        self._add_accumulator(self._moment1_acc_str, p, dtype=acc_dtype)
        self._add_accumulator(self._moment2_acc_str, p, dtype=acc_dtype)
        self._add_accumulator(
            name=self._beta1_pow_acc_str,
            param=p,
            dtype=acc_dtype,
            fill_value=(
                0.9 if isinstance(self._beta1, Variable) else self._beta1
            ),
            shape=[1],
            type=core.VarDesc.VarType.DENSE_TENSOR,
            device='cpu',
        )
        self._add_accumulator(
            name=self._beta2_pow_acc_str,
            param=p,
            dtype=acc_dtype,
            fill_value=(
                0.999 if isinstance(self._beta2, Variable) else self._beta2
            ),
            shape=[1],
            type=core.VarDesc.VarType.DENSE_TENSOR,
            device='cpu',
        )

    def _append_optimize_op(self, block, param_and_grad):
        if not isinstance(block, (framework.Block, pir.Block)):
            raise TypeError("block is not instance of Block.")
        if isinstance(param_and_grad, dict):
            param_and_grad = self._update_param_group(param_and_grad)

        block.program._use_lamb = True

        moment1 = self._get_accumulator_master(
            self._moment1_acc_str, param_and_grad[0]
        )
        moment2 = self._get_accumulator_master(
            self._moment2_acc_str, param_and_grad[0]
        )
        beta1_pow_acc = self._get_accumulator_master(
            self._beta1_pow_acc_str, param_and_grad[0]
        )
        beta2_pow_acc = self._get_accumulator_master(
            self._beta2_pow_acc_str, param_and_grad[0]
        )

        if (
            self._exclude_from_weight_decay_fn is not None
            and self._exclude_from_weight_decay_fn(param_and_grad[0])
        ):
            weight_decay = 0.0
        else:
            weight_decay = self._lamb_weight_decay
        lr = self._create_param_lr(param_and_grad)

        find_master = self._multi_precision and self._is_dtype_fp16_or_bf16(
            param_and_grad[0].dtype
        )
        p_name = param_and_grad[0].name
        if find_master:
            master_weight = self._master_weights[p_name]
            self._used_master_weights[p_name] = master_weight.name
        else:
            master_weight = None

        if framework.in_dynamic_or_pir_mode():
            _C_ops.lamb_(
                param_and_grad[0],
                param_and_grad[1],
                lr,
                moment1,
                moment2,
                beta1_pow_acc,
                beta2_pow_acc,
                master_weight,
                None,
                weight_decay,
                self._beta1,
                self._beta2,
                self._epsilon,
                self.always_adapt,
                find_master,
            )
            return None
        else:
            # create the lamb optimize op
            inputs = {
                "Param": param_and_grad[0],
                "Grad": param_and_grad[1],
                "LearningRate": lr,
                "Moment1": moment1,
                "Moment2": moment2,
                "Beta1Pow": beta1_pow_acc,
                "Beta2Pow": beta2_pow_acc,
            }
            outputs = {
                "ParamOut": param_and_grad[0],
                "Moment1Out": moment1,
                "Moment2Out": moment2,
                "Beta1PowOut": beta1_pow_acc,
                "Beta2PowOut": beta2_pow_acc,
            }
            attrs = {
                "beta1": self._beta1,
                "beta2": self._beta2,
                "epsilon": self._epsilon,
                "weight_decay": weight_decay,
                "always_adapt": self.always_adapt,
                "multi_precision": find_master,
            }

            if find_master:
                inputs["MasterParam"] = master_weight
                outputs["MasterParamOut"] = master_weight

            found_inf = self._get_auxiliary_var('found_inf')
            if found_inf:
                inputs["SkipUpdate"] = found_inf

            lamb_op = block.append_op(
                type=self.type,
                inputs=inputs,
                outputs=outputs,
                attrs=attrs,
                stop_gradient=True,
            )

            return lamb_op

    def _update_param_group(self, parameters):
        self._beta1 = parameters.get('beta1', self._default_dict['beta1'])
        self._beta2 = parameters.get('beta2', self._default_dict['beta2'])
        self._epsilon = parameters.get('epsilon', self._default_dict['epsilon'])
        self._lamb_weight_decay = parameters.get(
            'lamb_weight_decay', self._default_dict['lamb_weight_decay']
        )
        self._exclude_from_weight_decay_fn = parameters.get(
            'exclude_from_weight_decay_fn',
            self._default_dict['exclude_from_weight_decay_fn'],
        )
        parameters = parameters.get('params')
        return parameters