详解如何使用Pytorch进行多卡训练

import torch.nn as nn
import torch.optim as optim
import torch.utils.data as data
import torch.distributed as dist
import torch.multiprocessing as mp
# 定义网络模型
class Net(nn.Module):
    def __init__(self):
        super(Net, self).__init__()
        self.conv1 = nn.Conv2d(3, 32, kernel_size=5)
        self.relu = nn.ReLU()
        self.pool = nn.MaxPool2d(kernel_size=2, stride=2)
        self.fc1 = nn.Linear(4608, 64)
        self.fc2 = nn.Linear(64, 10)
    def forward(self, x):
        x = self.conv1(x)
        x = self.relu(x)
        x = self.pool(x)
        x = x.view(-1, 4608)
        x = self.fc1(x)
        x = self.relu(x)
        x = self.fc2(x)
        return x
# 定义训练函数
def train(gpu, args):
    rank = gpu
    dist.init_process_group(backend='nccl', init_method='env://', world_size=args.world_size, rank=rank)
    torch.cuda.set_device(gpu)
    train_loader = data.DataLoader(...)
    model = Net()
    model = nn.parallel.DistributedDataParallel(model, device_ids=[gpu])
    criterion = nn.CrossEntropyLoss()
    optimizer = optim.SGD(model.parameters(), lr=0.01)
    for epoch in range(args.epochs):
        epoch_loss = 0.0
        for i, (inputs, labels) in enumerate(train_loader):
            inputs, labels = inputs.cuda(), labels.cuda()
            optimizer.zero_grad()
            outputs = model(inputs)
            loss = criterion(outputs, labels)
            loss.backward()
            optimizer.step()
            epoch_loss += loss.item()
        print('GPU %d Loss: %.3f' % (gpu, epoch_loss))
# 主函数
if __name__ == '__main__':
    mp.set_start_method('spawn')
    args = parser.parse_args()
    args.world_size = args.num_gpus * args.nodes
    mp.spawn(train, args=(args,), nprocs=args.num_gpus, join=True)

import torch.nn as nn
import torch.optim as optim
import torch.multiprocessing as mp
import torch.distributed as dist
# 定义模型段
class SubNet(nn.Module):
    def __init__(self, in_features, out_features):
        super(SubNet, self).__init__()
        self.linear = nn.Linear(in_features, out_features)
    def forward(self, x):
        return self.linear(x)
# 定义整个模型
class Net(nn.Module):
    def __init__(self):
        super(Net, self).__init__()
        self.subnets = nn.ModuleList([
            SubNet(1024, 512),
            SubNet(512, 256),
            SubNet(256, 100)
        ])
    def forward(self, x):
        for subnet in self.subnets:
            x = subnet(x)
        return x
# 定义训练函数
def train(subnet_id, args):
    dist.init_process_group(backend='nccl', init_method='env://', world_size=args.world_size, rank=subnet_id)
    torch.cuda.set_device(subnet_id)
    train_loader = data.DataLoader(...)
    model = Net().cuda()
    criterion = nn.CrossEntropyLoss()
    optimizer = optim.SGD(model.parameters(), lr=0.01)
    for epoch in range(args.epochs):
        epoch_loss = 0.0
        for i, (inputs, labels) in enumerate(train_loader):
            inputs, labels = inputs.cuda(), labels.cuda()
            optimizer.zero_grad()
            outputs = model(inputs)
            loss = criterion(outputs, labels)
            loss.backward(retain_graph=True)  # 梯度保留，用于后续误差传播
            optimizer.step()
            epoch_loss += loss.item()
        if subnet_id == 0:
            print('Epoch %d Loss: %.3f' % (epoch, epoch_loss))
# 主函数
if __name__ == '__main__':
    mp.set_start_method('spawn')
    args = parser.parse_args()
    args.world_size = args.num_gpus * args.subnets
    tasks = []
    for i in range(args.subnets):
        tasks.append(mp.Process(target=train, args=(i, args)))
    for task in tasks:
        task.start()
    for task in tasks:
        task.join()