深度学习中的并行策略概述:2 Data Parallelism
深度学习中的并行策略概述:2 Data Parallelism
数据并行(Data Parallelism)的核心在于将模型的数据处理过程并行化。具体来说,面对大规模数据批次时,将其拆分为较小的子批次,并在多个计算设备上同时进行处理。每个设备负责处理一个子批次,实现并行计算。处理完成后,将各个设备上的计算结果汇总,以便对模型进行统一更新。由于其在深度学习中的普遍应用,数据并行成为了一种广泛支持的并行计算策略,并在主流框架中得到了良好的实现。
以下代码展示了如何在PyTorch中使用nn.DataParallel和DistributedDataParallel实现数据并行,以加速模型的训练过程。
使用nn.DataParallel实现数据并行
import torch
import torch.nn as nn
from torch.utils.data import Dataset, DataLoader
# 假设我们有一个简单的数据集类
class SimpleDataset(Dataset):
def __init__(self, data, target):
self.data = data
self.target = target
def __len__(self):
return len(self.data)
def __getitem__(self, idx):
return self.data[idx], self.target[idx]
# 假设我们有一个简单的神经网络模型
class SimpleModel(nn.Module):
def __init__(self, input_dim):
super(SimpleModel, self).__init__()
self.fc = nn.Linear(input_dim, 1)
def forward(self, x):
return torch.sigmoid(self.fc(x))
# 假设我们有一些数据
n_sample = 100
n_dim = 10
batch_size = 10
X = torch.randn(n_sample, n_dim)
Y = torch.randint(0, 2, (n_sample,)).float()
dataset = SimpleDataset(X, Y)
data_loader = DataLoader(dataset, batch_size=batch_size, shuffle=True)
# 初始化模型
device_ids = [0, 1, 2] # 指定使用的GPU编号
model = SimpleModel(n_dim).to(device_ids[0])
model = nn.DataParallel(model, device_ids=device_ids)
# 定义优化器和损失函数
optimizer = torch.optim.SGD(model.parameters(), lr=0.01)
criterion = nn.BCELoss()
# 训练模型
for epoch in range(10):
for batch_idx, (inputs, targets) in enumerate(data_loader):
inputs, targets = inputs.to('cuda'), targets.to('cuda')
outputs = model(inputs)
loss = criterion(outputs, targets.unsqueeze(1))
optimizer.zero_grad()
loss.backward()
optimizer.step()
print(f'Epoch {epoch}, Batch {batch_idx}, Loss: {loss.item()}')
使用DistributedDataParallel实现数据并行
import torch
import torch.nn as nn
import torch.optim as optim
from torch.utils.data import DataLoader, Dataset
import torch.distributed as dist
from torch.nn.parallel import DistributedDataParallel as DDP
# 假设我们有一个简单的数据集类
class SimpleDataset(Dataset):
def __init__(self, data, target):
self.data = data
self.target = target
def __len__(self):
return len(self.data)
def __getitem__(self, idx):
return self.data[idx], self.target[idx]
# 假设我们有一个简单的神经网络模型
class SimpleModel(nn.Module):
def __init__(self, input_dim):
super(SimpleModel, self).__init__()
self.fc = nn.Linear(input_dim, 1)
def forward(self, x):
return torch.sigmoid(self.fc(x))
# 初始化进程组
def init_process(rank, world_size, backend='nccl'):
dist.init_process_group(backend, rank=rank, world_size=world_size)
# 训练函数
def train(rank, world_size):
init_process(rank, world_size)
torch.cuda.set_device(rank)
model = SimpleModel(10).to(rank)
model = DDP(model, device_ids=[rank])
dataset = SimpleDataset(torch.randn(100, 10), torch.randint(0, 2, (100,)).float())
sampler = torch.utils.data.distributed.DistributedSampler(dataset, num_replicas=world_size, rank=rank)
data_loader = DataLoader(dataset, batch_size=10, sampler=sampler)
optimizer = optim.SGD(model.parameters(), lr=0.01)
criterion = nn.BCELoss()
for epoch in range(10):
for inputs, targets in data_loader:
inputs, targets = inputs.to(rank), targets.to(rank)
optimizer.zero_grad()
outputs = model(inputs)
loss = criterion(outputs, targets.unsqueeze(1))
loss.backward()
optimizer.step()
if __name__ == "__main__":
world_size = 4
torch.multiprocessing.spawn(train, args=(world_size,), nprocs=world_size, join=True)
原文地址:https://blog.csdn.net/shanglianlm/article/details/144682805
免责声明:本站文章内容转载自网络资源,如本站内容侵犯了原著者的合法权益,可联系本站删除。更多内容请关注自学内容网(zxcms.com)!