Training Process Exit Gracefully

Overview

When there are suboptimal devices in the training cluster, saving checkpoint and exiting the cluster training process before the failure occurs can effectively prevent the loss of weight data when the cluster is damaged. This also avoids issues such as training data rollback and loading checkpoint rollback when training recovery, effectively preventing the waste of training resources.

This document is an example of the training process exit gracefully. To illustrate the specific usage, we assume that the exit configuration detected at the first training step, and the training process is ended in advance. You can get the full sample code here: process_graceful_exit .

graceful_exit.py is the source code, train.sh is the start training script, and graceful_exit.json is the graceful exit config json file.

Dataset And Training Model

Data Preparation

Download the MNIST dataset and unzip the dataset to the project directory.

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

Model Definition

import os
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)

# dataset
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


# define the traning model
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

Environment Variable And Callback Function

Environment Variable

Using Training process Graceful Exit requires setting the environment variable MS_ENABLE_GRACEFUL_EXIT to 1. This environment variable can control the synchronization operator into the graph to ensure that all training processes can exit synchronously.

export MS_ENABLE_GRACEFUL_EXIT=1

Callback Function

In addition to the above of environment variable, it also needs to configure the callback function OnRequestExit , and passes the parameter config_file to provide the path of the graceful exit json file. This callback function will check if there is a graceful exit json file in the specified path at every training step begin. If the file exists, and the GracefulExit is 1 , it will save checkpoint and exit training process at current step end.

The GracefulExit in the Json file is dynamically configured during training. Generally, the keyword is modified when suboptimal devices exist in the training cluster and the training process needs to exit.

# key in json file：‘{“GracefulExit”: 1}’
config_json = r"./graceful_exit.json"

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

When configuring the OnRequestExit callback function, you can configure saving mindir, saving checkpoint, and other configuration parameters as required. For more details, please refer to the documentation OnRequestExit .

def graceful_exit_case():
    # initialize
    device_num = 8
    context.set_context(mode=context.GRAPH_MODE, device_target="Ascend")
    context.set_auto_parallel_context(parallel_mode=ParallelMode.SEMI_AUTO_PARALLEL, device_num=device_num)
    init()

    # model building
    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()})

    # the dependency file `reset.json`, like `{"GracefulExit": 1}`
    reset_json = r"./graceful_exit.json"

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

Starting Training

Using msrun to start training.

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

Analyzing The Results

After training ends, the following WARNING log will be printed: Graceful exit is triggered, stop training. Eight directories named rank_0 to rank_7 will be generated in the current execution directory, each containing a LeNet_train.ckpt file (if saving checkpoints is set in OnRequestExit ).

./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

Notes

If TrainOneStepCell is not overridden, you only need to configure the MS_ENABLE_GRACEFUL_EXIT environment variable, the OnRequestExit callback function, and modify the graceful exit json file as needed at a certain point during training.

If the network model requires overriding TrainOneStepCell:

The new method inherits from TrainOneStepCell , and the following if conditional branch code is added in the construct method to ensure the graceful exit feature works properly.

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))
        ...

The new method is not inherits from TrainOneStepCell , you need add the following code in __init__ method(don't change parameter's name), and using in the construct method.

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))
        ...