Amazon Web Services 文档中描述的 Amazon Web Services 服务或功能可能因区域而异。要查看适用于中国区域的差异，请参阅 中国的 Amazon Web Services 服务入门 (PDF)。

本文属于机器翻译版本。若本译文内容与英语原文存在差异，则一律以英文原文为准。

修改 PyTorch 训练脚本

在本节中，您将学习如何修改 PyTorch 训练脚本以配置用于自动分区和手动分区的 SageMaker 模型并行度库。

请注意，默认情况下启用自动分区。除非另行指定，否则以下脚本使用自动分区。

使用自动拆分 PyTorch

要使用模型并行度库运行 PyTorch 训练脚本，需要进行 SageMaker以下训练脚本更改：

要了解有关模型并行度库 API SageMaker 的更多信息，请参阅 AP I 文档。

import torch
import torch.nn as nn
import torch.nn.functional as F
import torch.optim as optim
from torchnet.dataset import SplitDataset
from torchvision import datasets

import smdistributed.modelparallel.torch as smp

class GroupedNet(nn.Module):
    def __init__(self):
        super(GroupedNet, self).__init__()
        # define layers

    def forward(self, x):
        # define forward pass and return model outputs


# smdistributed: Define smp.step. Return any tensors needed outside.
@smp.step
def train_step(model, data, target):
    output = model(data)
    loss = F.nll_loss(output, target, reduction="mean")
    model.backward(loss)
    return output, loss


def train(model, device, train_loader, optimizer):
    model.train()
    for batch_idx, (data, target) in enumerate(train_loader):
        # smdistributed: Move input tensors to the GPU ID used by the current process,
        # based on the set_device call.
        data, target = data.to(device), target.to(device)
        optimizer.zero_grad()
        # Return value, loss_mb is a StepOutput object
        _, loss_mb = train_step(model, data, target)

        # smdistributed: Average the loss across microbatches.
        loss = loss_mb.reduce_mean()

        optimizer.step()

# smdistributed: initialize the backend
smp.init()

# smdistributed: Set the device to the GPU ID used by the current process.
# Input tensors should be transferred to this device.
torch.cuda.set_device(smp.local_rank())
device = torch.device("cuda")

# smdistributed: Download only on a single process per instance.
# When this is not present, the file is corrupted by multiple processes trying
# to download and extract at the same time
dataset = datasets.MNIST("../data", train=True, download=False)

# smdistributed: Shard the dataset based on data-parallel ranks
if smp.dp_size() > 1:
    partitions_dict = {f"{i}": 1 / smp.dp_size() for i in range(smp.dp_size())}
    dataset = SplitDataset(dataset, partitions=partitions_dict)
    dataset.select(f"{smp.dp_rank()}")

# smdistributed: Set drop_last=True to ensure that batch size is always divisible
# by the number of microbatches
train_loader = torch.utils.data.DataLoader(dataset, batch_size=64, drop_last=True)

model = GroupedNet()
optimizer = optim.Adadelta(model.parameters(), lr=4.0)

# smdistributed: Use the DistributedModel container to provide the model
# to be partitioned across different ranks. For the rest of the script,
# the returned DistributedModel object should be used in place of
# the model provided for DistributedModel class instantiation.
model = smp.DistributedModel(model)
optimizer = smp.DistributedOptimizer(optimizer)

train(model, device, train_loader, optimizer)

使用手动拆分 PyTorch

使用 smp.partition 上下文管理器将模块放置在特定设备中。未放在任何 smp.partition 上下文中的任何模块都放在 default_partition 中。如果 auto_partition 设置为 False，则需要提供 default_partition。在特定 smp.partition 上下文中创建的模块将放置在对应的分区上。

要了解有关模型并行度库 API SageMaker 的更多信息，请参阅 AP I 文档。

import torch
import torch.nn as nn
import torch.nn.functional as F
import torch.optim as optim
from torchnet.dataset import SplitDataset
from torchvision import datasets

import smdistributed.modelparallel.torch as smp

class GroupedNet(nn.Module):
    def __init__(self):
        super(GroupedNet, self).__init__()
        with smp.partition(0):
            # define child modules on device 0
        with smp.partition(1):
            # define child modules on device 1

    def forward(self, x):
        # define forward pass and return model outputs


# smdistributed: Define smp.step. Return any tensors needed outside.
@smp.step
def train_step(model, data, target):
    output = model(data)
    loss = F.nll_loss(output, target, reduction="mean")
    model.backward(loss)
    return output, loss


def train(model, device, train_loader, optimizer):
    model.train()
    for batch_idx, (data, target) in enumerate(train_loader):
        # smdistributed: Move input tensors to the GPU ID used by the current process,
        # based on the set_device call.
        data, target = data.to(device), target.to(device)
        optimizer.zero_grad()
        # Return value, loss_mb is a StepOutput object
        _, loss_mb = train_step(model, data, target)

        # smdistributed: Average the loss across microbatches.
        loss = loss_mb.reduce_mean()

        optimizer.step()

# smdistributed: initialize the backend
smp.init()

# smdistributed: Set the device to the GPU ID used by the current process.
# Input tensors should be transferred to this device.
torch.cuda.set_device(smp.local_rank())
device = torch.device("cuda")

# smdistributed: Download only on a single process per instance.
# When this is not present, the file is corrupted by multiple processes trying
# to download and extract at the same time
dataset = datasets.MNIST("../data", train=True, download=False)

# smdistributed: Shard the dataset based on data-parallel ranks
if smp.dp_size() > 1:
    partitions_dict = {f"{i}": 1 / smp.dp_size() for i in range(smp.dp_size())}
    dataset = SplitDataset(dataset, partitions=partitions_dict)
    dataset.select(f"{smp.dp_rank()}")

# smdistributed: Set drop_last=True to ensure that batch size is always divisible
# by the number of microbatches
train_loader = torch.utils.data.DataLoader(dataset, batch_size=64, drop_last=True)

model = GroupedNet()
optimizer = optim.Adadelta(model.parameters(), lr=4.0)

# smdistributed: Use the DistributedModel container to provide the model
# to be partitioned across different ranks. For the rest of the script,
# the returned DistributedModel object should be used in place of
# the model provided for DistributedModel class instantiation.
model = smp.DistributedModel(model)
optimizer = smp.DistributedOptimizer(optimizer)

train(model, device, train_loader, optimizer)

注意事项

使用 SageMaker模型并行度库配置 PyTorch 训练脚本时，应注意以下几点：

不支持的框架功能

的模型并行度库不支持以下 PyTorch 功能： SageMaker