进程优雅退出

概述

当训练集群中存在亚健康设备时，如果能在亚健康设备发生故障之前完成 checkpoint 保存并结束集群训练进程，可以有效避免集群损坏时的权重数据丢失问题。同时，这也可以避免训练恢复时的数据回滚和 checkpoint 加载回滚等问题，从而减少训练资源的浪费。

本文档介绍进程优雅退出功能的使用方法。为了说明具体使用方式，以在第一个训练step时检测到退出配置信息并提前结束训练进程为例。您可以在这里下载完整代码：process_graceful_exit。

其中，graceful_exit.py 为训练脚本，train.sh 为 msrun 启动脚本，graceful_exit.json 为优雅退出配置文件。

数据和模型准备

准备数据

下载MNIST数据集，并解压数据集到项目目录。

wget http://mindspore-website.obs.cn-north-4.myhuaweicloud.com/notebook/datasets/MNIST_Data.zip
unzip MNIST_Data.zip

模型定义

import os
import mindspore as ms
import mindspore.context as context
import mindspore.dataset as ds
import mindspore.dataset.transforms as C
import mindspore.dataset.vision as CV
import mindspore.nn as nn
from mindspore.common import dtype as mstype
from mindspore.dataset.vision import Inter
from mindspore.train import Accuracy
from mindspore.train import Model, LossMonitor
from mindspore.train.callback import OnRequestExit
from mindspore.common.initializer import TruncatedNormal
from mindspore.communication.management import init
from mindspore.context import ParallelMode

context.set_context(mode=context.GRAPH_MODE)

# 数据集
DATASET_PATH = "./MNIST_Data"


def create_dataset(data_path, batch_size=32, repeat_size=1,
                   num_parallel_workers=1):
    """
    create dataset for train or test
    """
    # define dataset
    mnist_ds = ds.MnistDataset(data_path)

    resize_height, resize_width = 32, 32
    rescale = 1.0 / 255.0
    shift = 0.0
    rescale_nml = 1 / 0.3081
    shift_nml = -1 * 0.1307 / 0.3081

    # define map operations
    resize_op = CV.Resize((resize_height, resize_width), interpolation=Inter.LINEAR)  # Bilinear mode
    rescale_nml_op = CV.Rescale(rescale_nml, shift_nml)
    rescale_op = CV.Rescale(rescale, shift)
    hwc2chw_op = CV.HWC2CHW()
    type_cast_op = C.TypeCast(mstype.int32)

    # apply map operations on images
    mnist_ds = mnist_ds.map(operations=type_cast_op, input_columns="label", num_parallel_workers=num_parallel_workers)
    mnist_ds = mnist_ds.map(operations=resize_op, input_columns="image", num_parallel_workers=num_parallel_workers)
    mnist_ds = mnist_ds.map(operations=rescale_op, input_columns="image", num_parallel_workers=num_parallel_workers)
    mnist_ds = mnist_ds.map(operations=rescale_nml_op, input_columns="image", num_parallel_workers=num_parallel_workers)
    mnist_ds = mnist_ds.map(operations=hwc2chw_op, input_columns="image", num_parallel_workers=num_parallel_workers)

    # apply DatasetOps
    buffer_size = 10000
    mnist_ds = mnist_ds.shuffle(buffer_size=buffer_size)  # 10000 as in LeNet train script
    mnist_ds = mnist_ds.batch(batch_size, drop_remainder=True)
    mnist_ds = mnist_ds.repeat(repeat_size)

    return mnist_ds


# 定义网络模型
def conv(in_channels, out_channels, kernel_size, stride=1, padding=0):
    """weight initial for conv layer"""
    weight = weight_variable()
    return nn.Conv2d(in_channels, out_channels,
                     kernel_size=kernel_size, stride=stride, padding=padding,
                     weight_init=weight, has_bias=False, pad_mode="valid")


def fc_with_initialize(input_channels, out_channels):
    """weight initial for fc layer"""
    weight = weight_variable()
    bias = weight_variable()
    return nn.Dense(input_channels, out_channels, weight, bias)


def weight_variable():
    """weight initial"""
    return TruncatedNormal(0.02)


class LeNet5(nn.Cell):
    def __init__(self, num_class=10, channel=1):
        super(LeNet5, self).__init__()
        self.num_class = num_class
        self.conv1 = conv(channel, 6, 5)
        self.conv2 = conv(6, 16, 5)
        self.fc1 = fc_with_initialize(16 * 5 * 5, 120)
        self.fc2 = fc_with_initialize(120, 84)
        self.fc3 = fc_with_initialize(84, self.num_class)
        self.relu = nn.ReLU()
        self.max_pool2d = nn.MaxPool2d(kernel_size=2, stride=2)
        self.flatten = nn.Flatten()

    def construct(self, x):
        x = self.conv1(x)
        x = self.relu(x)
        x = self.max_pool2d(x)
        x = self.conv2(x)
        x = self.relu(x)
        x = self.max_pool2d(x)
        x = self.flatten(x)
        x = self.fc1(x)
        x = self.relu(x)
        x = self.fc2(x)
        x = self.relu(x)
        x = self.fc3(x)
        return x

环境变量以及Callback函数

环境变量

开启进程优雅退出功能，需要设置环境变量 MS_ENABLE_GRACEFUL_EXIT=1。该环境变量可以控制同步算子入图，来保证所有的训练进程可以同步退出。

export MS_ENABLE_GRACEFUL_EXIT=1

Callback函数

除了设置上述环境变量外，还需要设置 OnRequestExit Callback 函数，并通过参数 config_file 指定优雅退出配置文件的路径。该函数会在训练进程的每个 step 开始时检查指定目录下是否存在优雅退出配置文件。如果存在配置文件且文件中的关键字 GracefulExit 值为 1，则会在 step 结束时保存 checkpoint 文件并退出训练进程。

配置文件中的关键字 GracefulExit 是在训练过程中动态配置的，通常在识别到训练集群中存在亚健康设备且需要退出训练进程时进行修改。

# json文件中关键字：{"GracefulExit": 1}
config_json = r"./graceful_exit.json"

# 设置callback函数
cb = OnRequestExit(file_name="LeNet", config_file=config_json)

另外，在配置 OnRequestExit callback函数时，保存mindir、保存checkpoint以及其他配置参数可以根据需要自行配置，详情参见OnRequestExit 。

def graceful_exit_case():
    # init
    device_num = 8
    context.set_context(mode=context.GRAPH_MODE)
    ms.set_device("Ascend")

    context.set_auto_parallel_context(parallel_mode=ParallelMode.SEMI_AUTO_PARALLEL, device_num=device_num)
    init()

    # build
    network = LeNet5(10)
    ds_train = create_dataset(os.path.join(DATASET_PATH, "train"), 32, 1)
    net_loss = nn.SoftmaxCrossEntropyWithLogits(sparse=True, reduction="mean")
    net_opt = nn.Momentum(network.trainable_params(), 0.01, 0.9)
    model = Model(network, net_loss, net_opt, metrics={"Accuracy": Accuracy()})

    # graceful exit json file：{"GracefulExit": 1}
    reset_json = r"./graceful_exit.json"

    # callback
    cb = OnRequestExit(file_name="LeNet", config_file=reset_json)
    # train
    model.train(1, ds_train, callbacks=[cb, LossMonitor()], dataset_sink_mode=False)

启动训练

使用 msrun 命令启动训练。

msrun --worker_num=8 --local_worker_num=8 --master_addr=127.0.0.1 --master_port=10970 --join=True --log_dir=./comm_subgraph_logs graceful_exit_case.py

结果分析

训练结束后，日志中会有如下WARNING打印：Graceful exit is triggered, stop training 。同时，在当前执行目录下，会生成有 rank_0 至 rank_7 8个目录，每个目录下都有一个 LeNet_train.ckpt 文件（如果callback里面配置了保存checkpoint）。

./rank_0
├── LeNet_train.ckpt
└── LeNet_train.mindir
./rank_1
├── LeNet_train.ckpt
└── LeNet_train.mindir
./rank_2
├── LeNet_train.ckpt
└── LeNet_train.mindir
./rank_3
├── LeNet_train.ckpt
└── LeNet_train.mindir
./rank_4
├── LeNet_train.ckpt
└── LeNet_train.mindir
./rank_5
├── LeNet_train.ckpt
└── LeNet_train.mindir
./rank_6
├── LeNet_train.ckpt
└── LeNet_train.mindir
./rank_7
├── LeNet_train.ckpt
└── LeNet_train.mindir

说明

如果没有重写 TrainOneStepCell，则只需要配置 MS_ENABLE_GRACEFUL_EXIT 环境变量和 OnRequestExit callback 函数，并在训练的某一时刻按需修改优雅退出配置文件，即可实现进程优雅退出功能。

如果网络模型需要重写TrainOneStepCell，则：

继承父类TrainOneStepCell，construct方法里面添加如下 if 条件分支代码来保证优雅退出功能可以正常运行（继承于TrainOneStepCell，可以直接使用这些成员变量）：

class TrainOneStepCellWithABC(TrainOneStepCell):
    def __init__(self, ...):
        ...

    def construct(self, *inputs):
        ...
        grads = self.grad(self.network, self.weights)(*inputs, sens)
        if self.use_graceful_exit:
            grads = self.graceful_exit.exit_by_request(grads, self.init_param, self.exit_param)

        loss = F.depend(loss, self.optimizer(grads))
        ...

没有继承父类TrainOneStepCell，需要在 __init__ 方法里面新增如下代码（parameter的name不要修改），并在 construct 方法里面调用。示例代码如下：

from mindspore.utils import ExitByRequest

class TrainOneStepCellWithABC(Cell):
    def __init__(self, ...):
        ...
        self.use_graceful_exit = os.environ.get("MS_ENABLE_GRACEFUL_EXIT") == "1"
        if self.use_graceful_exit:
            self.graceful_exit = ExitByRequest()
            self.exit_param = Parameter(Tensor(False, mstype.bool_), name="graceful_exit")  # update by reduce value
            self.init_param = Parameter(Tensor([0], mstype.int32), name="graceful_init")  # update by config file

    def construct(self, *inputs):
        ...
        if self.use_graceful_exit:
            grads = self.graceful_exit.exit_by_request(grads, self.init_param, self.exit_param)
        loss = F.depend(loss, self.optimizer(grads))
        ...