utility.py

# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserve.
#
#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 absolute_import
from __future__ import division
from __future__ import print_function

import six
import argparse
import functools
import sys
import os
import logging
import warnings
import signal
import json

import numpy as np
import paddle
import paddle.fluid as fluid
from paddle.fluid.wrapped_decorator import signature_safe_contextmanager
from paddle.fluid.framework import Program, program_guard, name_scope, default_main_program
from paddle.fluid import unique_name, layers

import distutils.util
from utils import dist_utils

from utils.optimizer import Optimizer

logging.basicConfig(level=logging.INFO)
logger = logging.getLogger(__name__)


def print_arguments(args):
    """Print argparse's arguments.

    Usage:

    .. code-block:: python

        parser = argparse.ArgumentParser()
        parser.add_argument("name", default="Jonh", type=str, help="User name.")
        args = parser.parse_args()
        print_arguments(args)

    :param args: Input argparse.Namespace for printing.
    :type args: argparse.Namespace
    """

    logger.info("-------------  Configuration Arguments -------------")
    for arg, value in sorted(six.iteritems(vars(args))):
        logger.info("%25s : %s" % (arg, value))
    logger.info("----------------------------------------------------")


def add_arguments(argname, type, default, help, argparser, **kwargs):
    """Add argparse's argument.

    Usage:

    .. code-block:: python

        parser = argparse.ArgumentParser()
        add_argument("name", str, "Jonh", "User name.", parser)
        args = parser.parse_args()
    """
    type = distutils.util.strtobool if type == bool else type
    argparser.add_argument(
        "--" + argname,
        default=default,
        type=type,
        help=help + ' Default: %(default)s.',
        **kwargs)


def parse_args():
    """Add arguments

    Returns:
        all training args
    """
    parser = argparse.ArgumentParser(description=__doc__)
    add_arg = functools.partial(add_arguments, argparser=parser)
    # yapf: disable

    # ENV
    add_arg('use_gpu',                  bool,   True,                   "Whether to use GPU.")
    add_arg('model_save_dir',           str,    "./output",        "The directory path to save model.")
    add_arg('data_dir',                 str,    "./data/ILSVRC2012/",   "The ImageNet dataset root directory.")
    add_arg('pretrained_model',         str,    None,                   "Whether to load pretrained model.")
    add_arg('finetune_exclude_pretrained_params', str, None,            "Ignore params when doing finetune")
    add_arg('checkpoint',               str,    None,                   "Whether to resume checkpoint.")
    add_arg('print_step',               int,    10,                     "The steps interval to print logs")
    add_arg('save_step',                int,    1,                      "The steps interval to save checkpoints")

    # SOLVER AND HYPERPARAMETERS
    add_arg('model',                    str,    "ResNet50",   "The name of network.")
    add_arg('total_images',             int,    1281167,                "The number of total training images.")
    parser.add_argument('--image_shape', nargs='+', type=int, default=[3, 224, 224], help="The shape of image")
    add_arg('num_epochs',               int,    120,                    "The number of total epochs.")
    add_arg('class_dim',                int,    1000,                   "The number of total classes.")
    add_arg('batch_size',               int,    8,                      "Minibatch size on all the devices.")
    add_arg('test_batch_size',          int,    8,                   "Test batch size on all the devices.")
    add_arg('lr',                       float,  0.1,                    "The learning rate.")
    add_arg('lr_strategy',              str,    "piecewise_decay",      "The learning rate decay strategy.")
    add_arg('l2_decay',                 float,  1e-4,                   "The l2_decay parameter.")
    add_arg('momentum_rate',            float,  0.9,                    "The value of momentum_rate.")
    add_arg('warm_up_epochs',           float,  5.0,                    "The value of warm up epochs")
    add_arg('decay_epochs',             float,  2.4,                    "Decay epochs of exponential decay learning rate scheduler")
    add_arg('decay_rate',               float,  0.97,                   "Decay rate of exponential decay learning rate scheduler")
    add_arg('drop_connect_rate',        float,  0.2,                    "The value of drop connect rate")
    parser.add_argument('--step_epochs', nargs='+', type=int, default=[30, 60, 90], help="piecewise decay step")

    # READER AND PREPROCESS
    add_arg('use_dali',                 bool,   False,                  "Whether to use nvidia DALI for preprocessing")
    add_arg('lower_scale',              float,  0.08,                   "The value of lower_scale in ramdom_crop")
    add_arg('lower_ratio',              float,  3./4.,                  "The value of lower_ratio in ramdom_crop")
    add_arg('upper_ratio',              float,  4./3.,                  "The value of upper_ratio in ramdom_crop")
    add_arg('resize_short_size',        int,    256,                    "The value of resize_short_size")
    add_arg('use_mixup',                bool,   False,                  "Whether to use mixup")
    add_arg('mixup_alpha',              float,  0.2,                    "The value of mixup_alpha")
    add_arg('reader_thread',            int,    8,                      "The number of multi thread reader")
    add_arg('reader_buf_size',          int,    8,                      "The buf size of multi thread reader")
    add_arg('interpolation',            int,    None,                   "The interpolation mode")
    add_arg('use_aa',                   bool,   False,                  "Whether to use auto augment")
    parser.add_argument('--image_mean', nargs='+', type=float, default=[0.485, 0.456, 0.406], help="The mean of input image data")
    parser.add_argument('--image_std', nargs='+', type=float, default=[0.229, 0.224, 0.225], help="The std of input image data")

    # SWITCH
    add_arg('validate',                 bool,   True,                   "whether to validate when training.")
    add_arg('use_fp16',                 bool,   False,                  "Whether to enable half precision training with fp16." )
    add_arg('scale_loss',               float,  1.0,                    "The value of scale_loss for fp16." )
    add_arg('use_dynamic_loss_scaling', bool,   True,                   "Whether to use dynamic loss scaling.")
    add_arg('data_format',              str,    "NCHW",                 "Tensor data format when training.")
    add_arg('fuse_elewise_add_act_ops', bool,   False,                  "Whether to use elementwise_act fusion.")
    add_arg('fuse_bn_act_ops',          bool,   False,                  "Whether to use batch_norm and act fusion.")

    add_arg('use_label_smoothing',      bool,   False,                  "Whether to use label_smoothing")
    add_arg('label_smoothing_epsilon',  float,  0.1,                    "The value of label_smoothing_epsilon parameter")
    #NOTE: (2019/08/08) temporary disable use_distill
    #add_arg('use_distill',              bool,   False,                  "Whether to use distill")
    add_arg('use_ema',                  bool,   False,                  "Whether to use ExponentialMovingAverage.")
    add_arg('ema_decay',                float,  0.9999,                 "The value of ema decay rate")
    add_arg('padding_type',             str,    "SAME",                 "Padding type of convolution")
    add_arg('use_se',                   bool,   True,                   "Whether to use Squeeze-and-Excitation module for EfficientNet.")

    #NOTE: args for profiler
    add_arg("enable_ce",                bool,   False,                  "Whether to enable ce")
    add_arg('random_seed',              int,    None,                   "random seed")
    add_arg('is_profiler',              bool,   False,                  "Whether to start the profiler")
    add_arg('profiler_path',            str,    './profilier_files',                   "the profiler output file path")
    add_arg('max_iter',                 int,    0,                      "the max train batch num")
    add_arg('same_feed',                int,    0,                      "whether to feed same images")
    add_arg('enable_addto',             bool,   False,                  "whether to enable the addto strategy for gradient accumulation")
    add_arg('fuse_all_optimizer_ops',   bool,   False,                  "whether to fuse all optimizer operators")


    # yapf: enable
    args = parser.parse_args()

    return args


def check_gpu():
    """
    Log error and exit when set use_gpu=true in paddlepaddle
    cpu ver sion.
    """
    err = "Config use_gpu cannot be set as true while you are " \
                "using paddlepaddle cpu version ! \nPlease try: \n" \
                "\t1. Install paddlepaddle-gpu to run model on GPU \n" \
                "\t2. Set use_gpu as false in config file to run " \
                "model on CPU"
    try:
        if args.use_gpu and not fluid.is_compiled_with_cuda():
            logger.error(err)
            sys.exit(1)
    except Exception as e:
        pass


def check_version():
    """
    Log error and exit when the installed version of paddlepaddle is
    not satisfied.
    """
    err = "PaddlePaddle version 1.6 or higher is required, " \
          "or a suitable develop version is satisfied as well. \n" \
          "Please make sure the version is good with your code." \

    try:
        fluid.require_version('1.6.0')
    except Exception as e:
        logger.error(err)
        sys.exit(1)


def check_args(args):
    """check arguments before running

    Args:
        all arguments
    """

    # check models name
    sys.path.append("..")
    import models
    model_list = [m for m in dir(models) if "__" not in m]
    assert args.model in model_list, "{} is not in lists: {}, please check the model name".format(
        args.model, model_list)

    # check learning rate strategy
    lr_strategy_list = [l for l in dir(Optimizer) if not l.startswith('__')]
    if args.lr_strategy not in lr_strategy_list:
        logger.warning(
            "\n{} is not in lists: {}, \nUse default learning strategy now!".
            format(args.lr_strategy, lr_strategy_list))
        args.lr_strategy = "default_decay"

    # check confict of GoogLeNet and mixup
    if args.model == "GoogLeNet":
        assert args.use_mixup == False, "Cannot use mixup processing in GoogLeNet, please set use_mixup = False."

    # check interpolation of reader settings
    if args.interpolation:
        assert args.interpolation in [
            0, 1, 2, 3, 4
        ], "Wrong interpolation, please set:\n0: cv2.INTER_NEAREST\n1: cv2.INTER_LINEAR\n2: cv2.INTER_CUBIC\n3: cv2.INTER_AREA\n4: cv2.INTER_LANCZOS4"

    # check padding type
    if args.padding_type:
        assert args.padding_type in [
            "SAME", "VALID", "DYNAMIC"
        ], "Wrong padding_type, please set:\nSAME\nVALID\nDYNAMIC"

    # check checkpint and pretrained_model
    assert args.checkpoint is None or args.pretrained_model is None, "Do not init model by checkpoint and pretrained_model both."

    # check pretrained_model path for loading
    if args.pretrained_model is not None:
        assert isinstance(args.pretrained_model, str)
        assert os.path.isdir(
            args.
            pretrained_model), "please support available pretrained_model path."

    #FIXME: check checkpoint path for saving
    if args.checkpoint is not None:
        assert isinstance(args.checkpoint, str)
        assert os.path.isdir(
            args.checkpoint
        ), "please support available checkpoint path for initing model."

    # check gpu: when using gpu, the number of visible cards should divide batch size
    if args.use_gpu:
        assert args.batch_size % fluid.core.get_cuda_device_count(
        ) == 0, "please support correct batch_size({}), which can be divided by available cards({}), you can change the number of cards by indicating: export CUDA_VISIBLE_DEVICES= ".format(
            args.batch_size, fluid.core.get_cuda_device_count())

    # check data directory
    assert os.path.isdir(
        args.data_dir
    ), "Data doesn't exist in {}, please load right path".format(args.data_dir)

    # check CE
    if args.enable_ce:
        args.random_seed = 0
        logger.warning("CE is running now! already set random seed to 0")

    # check class_dim
    assert args.class_dim > 1, "class_dim must greater than 1"

    # check dali preprocess
    if args.use_dali:
        logger.warning(
            "DALI preprocessing is activated!!!\nWarning: 1. Please make sure paddlepaddle is compiled by GCC5.4 or later version!\n\t 2. Please make sure nightly builds DALI is installed correctly.\n----------------------------------------------------"
        )

    #check gpu
    check_gpu()
    check_version()


def init_model(exe, args, program):
    """load model from checkpoint or pretrained model
    """

    if args.checkpoint:
        fluid.io.load_persistables(exe, args.checkpoint, main_program=program)
        logger.info("Finish initing model from %s" % (args.checkpoint))

    if args.pretrained_model:
        """
        # yapf: disable
        # This is a dict of fc layers in all the classification models.
        final_fc_name = [
                         "fc8_weights","fc8_offset", #alexnet
                         "fc_weights","fc_offset", #darknet, densenet, dpn, hrnet, mobilenet_v3, res2net, res2net_vd, resnext, resnext_vd, xception
                         #efficient
                         "out","out_offset", "out1","out1_offset", "out2","out2_offset", #googlenet
                         "final_fc_weights", "final_fc_offset", #inception_v4
                         "fc7_weights", "fc7_offset", #mobilenetv1
                         "fc10_weights", "fc10_offset", #mobilenetv2
                         "fc_0", #resnet, resnet_vc, resnet_vd
                         "fc.weight", "fc.bias", #resnext101_wsl
                         "fc6_weights", "fc6_offset", #se_resnet_vd, se_resnext, se_resnext_vd, shufflenet_v2, shufflenet_v2_swish,
                         #squeezenet
                         "fc8_weights", "fc8_offset", #vgg
                         "fc_bias" #"fc_weights", xception_deeplab
                         ]
        # yapf: enable
        """
        final_fc_name = []
        if args.finetune_exclude_pretrained_params:
            final_fc_name = [
                str(s)
                for s in args.finetune_exclude_pretrained_params.split(",")
            ]

        def is_parameter(var):
            fc_exclude_flag = False
            for item in final_fc_name:
                if item in var.name:
                    fc_exclude_flag = True

            return isinstance(
                var, fluid.framework.
                Parameter) and not fc_exclude_flag and os.path.exists(
                    os.path.join(args.pretrained_model, var.name))

        logger.info("Load pretrain weights from {}, exclude params {}.".format(
            args.pretrained_model, final_fc_name))
        vars = filter(is_parameter, program.list_vars())
        fluid.io.load_vars(
            exe, args.pretrained_model, vars=vars, main_program=program)


def save_model(args, exe, train_prog, info):
    """save model in model_path
    """

    model_path = os.path.join(args.model_save_dir, args.model, str(info))
    if not os.path.isdir(model_path):
        os.makedirs(model_path)
    fluid.io.save_persistables(exe, model_path, main_program=train_prog)
    logger.info("Already save model in %s" % (model_path))


def save_json(info, path):
    """ save eval result or infer result to file as json format.
    """
    with open(path, 'w') as f:
        json.dump(info, f)


def create_data_loader(is_train, args):
    """create data_loader

    Usage:
        Using mixup process in training, it will return 5 results, include data_loader, image, y_a(label), y_b(label) and lamda, or it will return 3 results, include data_loader, image, and label.

    Args:
        is_train: mode
        args: arguments

    Returns:
        data_loader and the input data of net,
    """
    image_shape = args.image_shape
    feed_image = fluid.data(
        name="feed_image",
        shape=[None] + image_shape,
        dtype="float32",
        lod_level=0)

    feed_label = fluid.data(
        name="feed_label", shape=[None, 1], dtype="int64", lod_level=0)
    feed_y_a = fluid.data(
        name="feed_y_a", shape=[None, 1], dtype="int64", lod_level=0)

    capacity = 64 if int(os.environ.get('PADDLE_TRAINERS_NUM', 1)) <= 1 else 8

    if is_train and args.use_mixup:
        feed_y_b = fluid.data(
            name="feed_y_b", shape=[None, 1], dtype="int64", lod_level=0)
        feed_lam = fluid.data(
            name="feed_lam", shape=[None, 1], dtype="float32", lod_level=0)

        data_loader = fluid.io.DataLoader.from_generator(
            feed_list=[feed_image, feed_y_a, feed_y_b, feed_lam],
            capacity=capacity,
            use_double_buffer=True,
            iterable=True)
        return data_loader, [feed_image, feed_y_a, feed_y_b, feed_lam]
    else:
        if args.use_dali:
            return None, [feed_image, feed_label]

        data_loader = fluid.io.DataLoader.from_generator(
            feed_list=[feed_image, feed_label],
            capacity=capacity,
            use_double_buffer=True,
            iterable=True)

        return data_loader, [feed_image, feed_label]


def print_info(info_mode,
               metrics,
               time_info,
               pass_id=0,
               batch_id=0,
               print_step=1,
               device_num=1,
               class_dim=5):
    """print function

    Args:
        pass_id: epoch index
        batch_id: batch index
        print_step: the print_step arguments
        metrics: message to print
        time_info: time infomation
        info_mode: mode
    """
    #XXX: Use specific name to choose pattern, not the length of metrics.
    if info_mode == "batch":
        if batch_id % print_step == 0:
            #if isinstance(metrics,np.ndarray):
            # train and mixup output
            if len(metrics) == 2:
                loss, lr = metrics
                logger.info(
                    "[Pass {0}, train batch {1}] \tloss {2}, lr {3}, elapse {4}".
                    format(pass_id, batch_id, "%.5f" % loss, "%.5f" % lr,
                           "%2.4f sec" % time_info))
            # train and no mixup output
            elif len(metrics) == 4:
                loss, acc1, acc5, lr = metrics
                logger.info(
                    "[Pass {0}, train batch {1}] \tloss {2}, acc1 {3}, acc{7} {4}, lr {5}, elapse {6}".
                    format(pass_id, batch_id, "%.5f" % loss, "%.5f" % acc1,
                           "%.5f" % acc5, "%.5f" % lr, "%2.4f sec" % time_info,
                           min(class_dim, 5)))
            # test output
            elif len(metrics) == 3:
                loss, acc1, acc5 = metrics
                logger.info(
                    "[Pass {0}, test  batch {1}] \tloss {2}, acc1 {3}, acc{6} {4}, elapse {5}".
                    format(pass_id, batch_id, "%.5f" % loss, "%.5f" % acc1,
                           "%.5f" % acc5, "%2.4f sec" % time_info,
                           min(class_dim, 5)))
            else:
                raise Exception(
                    "length of metrics {} is not implemented, It maybe caused by wrong format of build_program_output".
                    format(len(metrics)))
            sys.stdout.flush()

    elif info_mode == "epoch":
        ## TODO add time elapse
        if len(metrics) == 5:
            train_loss, _, test_loss, test_acc1, test_acc5 = metrics
            logger.info(
                "[End pass {0}]\ttrain_loss {1}, test_loss {2}, test_acc1 {3}, test_acc{5} {4}".
                format(pass_id, "%.5f" % train_loss, "%.5f" % test_loss, "%.5f"
                       % test_acc1, "%.5f" % test_acc5, min(class_dim, 5)))
        elif len(metrics) == 7:
            train_loss, train_acc1, train_acc5, _, test_loss, test_acc1, test_acc5 = metrics
            logger.info(
                "[End pass {0}]\ttrain_loss {1}, train_acc1 {2}, train_acc{7} {3},test_loss {4}, test_acc1 {5}, test_acc{7} {6}".
                format(pass_id, "%.5f" % train_loss, "%.5f" % train_acc1, "%.5f"
                       % train_acc5, "%.5f" % test_loss, "%.5f" % test_acc1,
                       "%.5f" % test_acc5, min(class_dim, 5)))
        sys.stdout.flush()
    elif info_mode == "ce":
        assert len(
            metrics
        ) == 7, "Enable CE: The Metrics should contain train_loss, train_acc1, train_acc5, test_loss, test_acc1, test_acc5, and train_speed"
        assert len(
            time_info
        ) > 10, "0~9th batch statistics will drop when doing benchmark or ce, because it might be mixed with startup time, so please make sure training at least 10 batches."
        print_ce(device_num, metrics, time_info)
    else:
        raise Exception("Illegal info_mode")


def print_ce(device_num, metrics, time_info):
    """ Print log for CE(for internal test).
    """
    train_loss, train_acc1, train_acc5, _, test_loss, test_acc1, test_acc5 = metrics

    train_speed = np.mean(np.array(time_info[10:]))

    logger.info("kpis\ttrain_cost_card{}\t{}".format(device_num, train_loss))
    logger.info("kpis\ttrain_acc1_card{}\t{}".format(device_num, train_acc1))
    logger.info("kpis\ttrain_acc5_card{}\t{}".format(device_num, train_acc5))
    logger.info("kpis\ttest_cost_card{}\t{}".format(device_num, test_loss))
    logger.info("kpis\ttest_acc1_card{}\t{}".format(device_num, test_acc1))
    logger.info("kpis\ttest_acc5_card{}\t{}".format(device_num, test_acc5))
    logger.info("kpis\ttrain_speed_card{}\t{}".format(device_num, train_speed))


def best_strategy_compiled(args,
                           program,
                           loss,
                           exe,
                           mode="train",
                           share_prog=None):
    """make a program which wrapped by a compiled program
    """

    if os.getenv('FLAGS_use_ngraph'):
        return program
    else:
        build_strategy = fluid.compiler.BuildStrategy()
        try:
            fluid.require_version(min_version='1.7.0')
            build_strategy.fuse_bn_act_ops = args.fuse_bn_act_ops
            build_strategy.fuse_all_optimizer_ops = args.fuse_all_optimizer_ops
            build_strategy.enable_addto = args.enable_addto
        except Exception as e:
            logger.info("PaddlePaddle version 1.7.0 or higher is "
            "required when you want to fuse batch_norm and activation_op.")
        build_strategy.fuse_elewise_add_act_ops = args.fuse_elewise_add_act_ops

        exec_strategy = fluid.ExecutionStrategy()

        if args.use_gpu:
            exec_strategy.num_threads = fluid.core.get_cuda_device_count()

        exec_strategy.num_iteration_per_drop_scope = 10

        num_trainers = int(os.environ.get('PADDLE_TRAINERS_NUM', 1))
        if num_trainers > 1 and args.use_gpu:
            dist_utils.prepare_for_multi_process(exe, build_strategy, program)
            # NOTE: the process is fast when num_threads is 1
            # for multi-process training.
            exec_strategy.num_threads = 1

        compiled_program = fluid.CompiledProgram(program).with_data_parallel(
            loss_name=loss.name if mode == "train" else None,
            share_vars_from=share_prog if mode == "val" else None,
            build_strategy=build_strategy,
            exec_strategy=exec_strategy)

        return compiled_program


class ExponentialMovingAverage(object):
    def __init__(self,
                 decay=0.999,
                 thres_steps=None,
                 zero_debias=False,
                 name=None):
        self._decay = decay
        self._thres_steps = thres_steps
        self._name = name if name is not None else ''
        self._decay_var = self._get_ema_decay()

        self._params_tmps = []
        for param in default_main_program().global_block().all_parameters():
            if param.do_model_average != False:
                tmp = param.block.create_var(
                    name=unique_name.generate(".".join(
                        [self._name + param.name, 'ema_tmp'])),
                    dtype=param.dtype,
                    persistable=False,
                    stop_gradient=True)
                self._params_tmps.append((param, tmp))

        self._ema_vars = {}
        for param, tmp in self._params_tmps:
            with param.block.program._optimized_guard(
                [param, tmp]), name_scope('moving_average'):
                self._ema_vars[param.name] = self._create_ema_vars(param)

        self.apply_program = Program()
        block = self.apply_program.global_block()
        with program_guard(main_program=self.apply_program):
            decay_pow = self._get_decay_pow(block)
            for param, tmp in self._params_tmps:
                param = block._clone_variable(param)
                tmp = block._clone_variable(tmp)
                ema = block._clone_variable(self._ema_vars[param.name])
                layers.assign(input=param, output=tmp)
                # bias correction
                if zero_debias:
                    ema = ema / (1.0 - decay_pow)
                layers.assign(input=ema, output=param)

        self.restore_program = Program()
        block = self.restore_program.global_block()
        with program_guard(main_program=self.restore_program):
            for param, tmp in self._params_tmps:
                tmp = block._clone_variable(tmp)
                param = block._clone_variable(param)
                layers.assign(input=tmp, output=param)

    def _get_ema_decay(self):
        with default_main_program()._lr_schedule_guard():
            decay_var = layers.tensor.create_global_var(
                shape=[1],
                value=self._decay,
                dtype='float32',
                persistable=True,
                name="scheduled_ema_decay_rate")

            if self._thres_steps is not None:
                decay_t = (self._thres_steps + 1.0) / (self._thres_steps + 10.0)
                with layers.control_flow.Switch() as switch:
                    with switch.case(decay_t < self._decay):
                        layers.tensor.assign(decay_t, decay_var)
                    with switch.default():
                        layers.tensor.assign(
                            np.array(
                                [self._decay], dtype=np.float32),
                            decay_var)
        return decay_var

    def _get_decay_pow(self, block):
        global_steps = layers.learning_rate_scheduler._decay_step_counter()
        decay_var = block._clone_variable(self._decay_var)
        decay_pow_acc = layers.elementwise_pow(decay_var, global_steps + 1)
        return decay_pow_acc

    def _create_ema_vars(self, param):
        param_ema = layers.create_global_var(
            name=unique_name.generate(self._name + param.name + '_ema'),
            shape=param.shape,
            value=0.0,
            dtype=param.dtype,
            persistable=True)

        return param_ema

    def update(self):
        """
        Update Exponential Moving Average. Should only call this method in
        train program.
        """
        param_master_emas = []
        for param, tmp in self._params_tmps:
            with param.block.program._optimized_guard(
                [param, tmp]), name_scope('moving_average'):
                param_ema = self._ema_vars[param.name]
                if param.name + '.master' in self._ema_vars:
                    master_ema = self._ema_vars[param.name + '.master']
                    param_master_emas.append([param_ema, master_ema])
                else:
                    ema_t = param_ema * self._decay_var + param * (
                        1 - self._decay_var)
                    layers.assign(input=ema_t, output=param_ema)

        # for fp16 params
        for param_ema, master_ema in param_master_emas:
            default_main_program().global_block().append_op(
                type="cast",
                inputs={"X": master_ema},
                outputs={"Out": param_ema},
                attrs={
                    "in_dtype": master_ema.dtype,
                    "out_dtype": param_ema.dtype
                })

    @signature_safe_contextmanager
    def apply(self, executor, need_restore=True):
        """
        Apply moving average to parameters for evaluation.

        Args:
            executor (Executor): The Executor to execute applying.
            need_restore (bool): Whether to restore parameters after applying.
        """
        executor.run(self.apply_program)
        try:
            yield
        finally:
            if need_restore:
                self.restore(executor)

    def restore(self, executor):
        """Restore parameters.

        Args:
            executor (Executor): The Executor to execute restoring.
        """
        executor.run(self.restore_program)