adam.py

# Copyright (c) 2020 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 .optimizer import Optimizer
from ..fluid import core
from ..fluid import framework
from ..fluid.framework import Variable

import paddle
from paddle.fluid.dygraph.parallel import apply_collective_grads

__all__ = ["Adam"]


class Adam(Optimizer):
    r"""
    The Adam optimizer uses an optimization described at the end
    of section 2 of `Adam paper <https://arxiv.org/abs/1412.6980>`_ ,
    it can dynamically adjusts the learning rate of each parameter using
    the 1st moment estimates and the 2nd moment estimates of the gradient.

    The parameter ``param_out`` update rule with gradient ``grad``:

    .. math::

        t & = t + 1

        moment\_1\_out & = {\\beta}_1 * moment\_1 + (1 - {\\beta}_1) * grad

        moment\_2\_out & = {\\beta}_2 * moment\_2 + (1 - {\\beta}_2) * grad * grad

        learning\_rate & = learning\_rate * \\
                          \\frac{\sqrt{1 - {\\beta}_2^t}}{1 - {\\beta}_1^t}

        param\_out & = param - learning\_rate * \\frac{moment\_1}{\sqrt{moment\_2} + \epsilon}

    Related paper: `Adam: A Method for Stochastic Optimization <https://arxiv.org/abs/1412.6980>`_

    Args:
        learning_rate (float|LRScheduler, optional): The learning rate used to update ``Parameter``.
            It can be a float value or a LRScheduler. The default value is 0.001.
        beta1 (float|Tensor, optional): The exponential decay rate for the 1st moment estimates.
            It should be a float number or a Tensor with shape [1] and data type as float32.
            The default value is 0.9.
        beta2 (float|Tensor, optional): The exponential decay rate for the 2nd moment estimates.
            It should be a float number or a Tensor with shape [1] and data type as float32.
            The default value is 0.999.
        epsilon (float, optional): A small float value for numerical stability.
            The default value is 1e-08.
	parameters (list, optional): List of ``Tensor`` to update to minimize ``loss``. \
	    This parameter is required in dygraph mode. \
	    The default value is None in static mode, at this time all parameters will be updated.
	weight_decay (float|WeightDecayRegularizer, optional): The strategy of regularization. \
	    It canbe a float value as coeff of L2 regularization or \
	    :ref:`api_fluid_regularizer_L1Decay`, :ref:`api_fluid_regularizer_L2Decay`.
	    If a parameter has set regularizer using :ref:`api_fluid_ParamAttr` already, \
	    the regularization setting here in optimizer will be ignored for this parameter. \
	    Otherwise, the regularization setting here in optimizer will take effect. \
	    Default None, meaning there is no regularization.
        grad_clip (GradientClipBase, optional): Gradient cliping strategy, it's an instance of
            some derived class of ``GradientClipBase`` . There are three cliping strategies
            ( :ref:`api_fluid_clip_GradientClipByGlobalNorm` , :ref:`api_fluid_clip_GradientClipByNorm` ,
            :ref:`api_fluid_clip_GradientClipByValue` ). Default None, meaning there is no gradient clipping.
        lazy_mode (bool, optional): The official Adam algorithm has two moving-average accumulators.
            The accumulators are updated at every step. Every element of the two moving-average
            is updated in both dense mode and sparse mode. If the size of parameter is very large,
            then the update may be very slow. The lazy mode only update the element that has
            gradient in current mini-batch, so it will be much more faster. But this mode has
            different semantics with the original Adam algorithm and may lead to different result.
            The default value is False.
        name (str, optional): Normally there is no need for user to set this property.
            For more information, please refer to :ref:`api_guide_Name`.
            The default value is None.

    Examples:
        .. code-block:: python

            import paddle

            linear = paddle.nn.Linear(10, 10)
            inp = paddle.rand([10,10], dtype="float32")
            out = linear(inp)
            loss = paddle.mean(out)
            adam = paddle.optimizer.Adam(learning_rate=0.1,
                    parameters=linear.parameters())
            out.backward()
            adam.step()
            adam.clear_grad()

        .. code-block:: python

            # Adam with beta1/beta2 as Tensor and weight_decay as float
            import paddle

            linear = paddle.nn.Linear(10, 10)
            inp = paddle.rand([10,10], dtype="float32")
            out = linear(inp)
            loss = paddle.mean(out)

            beta1 = paddle.to_tensor([0.9], dtype="float32")
            beta2 = paddle.to_tensor([0.99], dtype="float32")

            adam = paddle.optimizer.Adam(learning_rate=0.1,
                    parameters=linear.parameters(),
                    beta1=beta1,
                    beta2=beta2,
                    weight_decay=0.01)
            out.backward()
            adam.step()
            adam.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=0.001,
                 beta1=0.9,
                 beta2=0.999,
                 epsilon=1e-8,
                 parameters=None,
                 weight_decay=None,
                 grad_clip=None,
                 lazy_mode=False,
                 name=None):
        assert learning_rate is not None
        assert beta1 is not None
        assert beta2 is not None
        assert epsilon is not None
        if not 0 <= beta1 < 1:
            raise ValueError("Invaild value of beta1, expect beta1 in [0,1).")
        if not 0 <= beta2 < 1:
            raise ValueError("Invaild value of beta2, expect beta2 in [0,1).")
        if not 0 <= epsilon:
            raise ValueError("Invaild value of epsilon, expect epsilon >= 0.")
        super(Adam, self).__init__(
            learning_rate=learning_rate,
            parameters=parameters,
            weight_decay=weight_decay,
            grad_clip=grad_clip,
            name=name)
        self.type = "adam"
        self._beta1 = beta1
        self._beta2 = beta2
        self._epsilon = epsilon
        self._lazy_mode = lazy_mode

    def _create_accumulators(self, block, parameters):
        assert isinstance(block, framework.Block)

        # Create accumulator tensors for first and second moments
        for p in parameters:
            self._add_accumulator(self._moment1_acc_str, p)
            self._add_accumulator(self._moment2_acc_str, p)
            self._add_accumulator(
                name=self._beta1_pow_acc_str,
                param=p,
                fill_value=0.9 if isinstance(self._beta1, Variable) \
                        else self._beta1,
                shape=[1],
                type=core.VarDesc.VarType.LOD_TENSOR, device='cpu')
            self._add_accumulator(
                name=self._beta2_pow_acc_str,
                param=p,
                fill_value=0.999 if isinstance(self._beta2, Variable) \
                        else self._beta2,
                shape=[1],
                type=core.VarDesc.VarType.LOD_TENSOR, device='cpu')

    def _append_optimize_op(self, block, param_and_grad):
        assert isinstance(block, framework.Block)

        moment1 = self._get_accumulator(self._moment1_acc_str,
                                        param_and_grad[0])
        moment2 = self._get_accumulator(self._moment2_acc_str,
                                        param_and_grad[0])
        beta1_pow_acc = self._get_accumulator(self._beta1_pow_acc_str,
                                              param_and_grad[0])
        beta2_pow_acc = self._get_accumulator(self._beta2_pow_acc_str,
                                              param_and_grad[0])
        lr = self._create_param_lr(param_and_grad)
        # create the adam optimize op

        if framework.in_dygraph_mode():
            _beta1 = self._beta1 if not isinstance(
                self._beta1, Variable) else self._beta1.numpy().item(0)
            _beta2 = self._beta2 if not isinstance(
                self._beta2, Variable) else self._beta2.numpy().item(0)
            _, _, _, _, _ = core.ops.adam(
                param_and_grad[0], param_and_grad[1], lr, moment1, moment2,
                beta1_pow_acc, beta2_pow_acc, param_and_grad[0], moment1,
                moment2, beta1_pow_acc, beta2_pow_acc, 'epsilon', self._epsilon,
                'lazy_mode', self._lazy_mode, 'min_row_size_to_use_multithread',
                1000, 'beta1', _beta1, 'beta2', _beta2)

            return None

        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 = {
            "epsilon": self._epsilon,
            "lazy_mode": self._lazy_mode,
            "min_row_size_to_use_multithread": 1000
        }

        if isinstance(self._beta1, Variable):
            inputs['Beta1Tensor'] = self._beta1
        else:
            attrs['beta1'] = self._beta1
        if isinstance(self._beta2, Variable):
            inputs['Beta2Tensor'] = self._beta2
        else:
            attrs['beta2'] = self._beta2

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

        return adam_op

    @framework.dygraph_only
    def step(self):
        """
        Execute the optimizer and update parameters once.

        Returns:
            None

        Examples:
            .. code-block:: python

                import paddle
                
                a = paddle.rand([2,13], dtype="float32")
                linear = paddle.nn.Linear(13, 5)
                # This can be any optimizer supported by dygraph.
                adam = paddle.optimizer.Adam(learning_rate = 0.01,
                                            parameters = linear.parameters())
                out = linear(a)
                out.backward()
                adam.step()
                adam.clear_grad()
        """
        if paddle.distributed.get_world_size() > 1:
            apply_collective_grads(self._parameter_list)

        self._dtype = None
        params_grads = []
        for param in self._parameter_list:
            if not param.trainable:
                continue
            if param._grad_ivar() is not None:
                grad_var = param._grad_ivar()
                if hasattr(grad_var, "_is_sparse") and grad_var._is_sparse(
                ) and self.regularization is not None:
                    raise RuntimeError(
                        "Adam don't support weight_decay with sparse parameters, please set it to None."
                    )
                params_grads.append((param, grad_var))

        optimize_ops = self._apply_optimize(
            loss=None, startup_program=None, params_grads=params_grads)