进行多GPU训练有几个主要原因:
提高训练速度: 使用多个GPU可以将模型的参数和训练数据分配到不同的GPU上并行处理,从而显著提高训练速度。每个GPU都可以处理一部分数据,同时进行反向传播和参数更新,使得整个训练过程更加高效。
处理更大的模型和数据集: 多GPU训练使得可以处理更大的模型和数据集,因为每个GPU都可以专注于处理部分模型参数和数据。这对于深度学习中复杂模型和大规模数据集的训练非常有益。
资源利用率提高: 利用多个GPU可以更充分地利用计算资源。在单个GPU上,可能存在计算资源的浪费,而多GPU训练可以更有效地利用这些资源。
实验迭代速度提高: 多GPU训练还有助于提高实验迭代速度。在深度学习中,通过更快地尝试不同的模型架构、超参数和数据处理方法,可以更快地找到最优的模型配置。
需要注意的是,多GPU训练也面临一些挑战,如数据同步、通信开销等问题,需要采用适当的并行训练策略来解决这些问题。
1. 从0实现
多GPU训练的核心不同的卡得到不同的样本个数,分别进行训练然后将得到的各自梯度进行相加
%matplotlib inline
import torch
from torch import nn
from torch.nn import functional as F
from d2l import torch as d2l
# 以LeNet网络为例子
scale = 0.01
W1 = torch.randn(size=(20, 1, 3, 3)) * scale
b1 = torch.zeros(20)
W2 = torch.randn(size=(50, 20, 5, 5)) * scale
b2 = torch.zeros(50)
W3 = torch.randn(size=(800, 128)) * scale
b3 = torch.zeros(128)
W4 = torch.randn(size=(128, 10)) * scale
b4 = torch.zeros(10)
params = [W1, b1, W2, b2, W3, b3, W4, b4]
def lenet(X, params):
h1_conv = F.conv2d(input=X, weight=params[0], bias=params[1])
h1_activation = F.relu(h1_conv)
h1 = F.avg_pool2d(input=h1_activation, kernel_size=(2, 2), stride=(2, 2))
h2_conv = F.conv2d(input=h1, weight=params[2], bias=params[3])
h2_activation = F.relu(h2_conv)
h2 = F.avg_pool2d(input=h2_activation, kernel_size=(2, 2), stride=(2, 2))
h2 = h2.reshape(h2.shape[0], -1)
h3_linear = torch.mm(h2, params[4]) + params[5]
h3 = F.relu(h3_linear)
y_hat = torch.mm(h3, params[6]) + params[7]
return y_hat
loss = nn.CrossEntropyLoss(reduction='none')
向多个设备分发参数
# 把一个参数复制到指定device的gpu上
def get_params(params, device): # params是list,用来存放各个参数
new_params = [p.clone().to(device) for p in params]
for p in new_params:
p.requires_grad_()
return new_params
new_params = get_params(params, d2l.try_gpu(0)) # 将params全部移动到cuda0上
print('b1 weight:', new_params[1])
print('b1 grad:', new_params[1].grad)
allreduce 函数将所有向量【将各个梯度的data相加然后复制回去】相加,并将结果广播给所有 GPU
def allreduce(data): # data是一个list[有几张卡就有几个list]
for i in range(1, len(data)): # 把从1到后面cuda的数据全部复制到cuda0上,因为加法计算必须在同一张卡上操作
data[0][:] += data[i].to(data[0].device) # for结束后, data[0中存放的就是所有梯度相加的结果
for i in range(1, len(data)):# 将data[0]中的数据复制回各个gpu
data[i] = data[0].to(data[i].device)
# test
data = [torch.ones((1, 2), device=d2l.try_gpu(i)) * (i + 1) for i in range(2)]
print('before allreduce:\n', data[0], '\n', data[1])
allreduce(data)
print('after allreduce:\n', data[0], '\n', data[1])
"""
output
before allreduce:
tensor([[1., 1.]], device='cuda:0')
tensor([[2., 2.]], device='cuda:1')
after allreduce:
tensor([[3., 3.]], device='cuda:0')
tensor([[3., 3.]], device='cuda:1')
"""
将一个小批量数据均匀地分布在多个 GPU 上
# test
data = torch.arange(20).reshape(4, 5)
devices = [torch.device('cuda:0'), torch.device('cuda:1')]
split = nn.parallel.scatter(data, devices) # .scatter直接均匀切割
print('input :', data)
print('load into', devices)
print('output:', split)
"""
output
input : tensor([[ 0, 1, 2, 3, 4],
[ 5, 6, 7, 8, 9],
[10, 11, 12, 13, 14],
[15, 16, 17, 18, 19]])
load into [device(type='cuda', index=0), device(type='cuda', index=1)]
output: (tensor([[0, 1, 2, 3, 4],
[5, 6, 7, 8, 9]], device='cuda:0'),
tensor([[10, 11, 12, 13, 14],
[15, 16, 17, 18, 19]], device='cuda:1'))
"""
def split_batch(X, y, devices):
"""将`X`和`y`拆分到多个设备上"""
assert X.shape[0] == y.shape[0]
return (nn.parallel.scatter(X, devices),
nn.parallel.scatter(y, devices))
在一个小批量上实现多 GPU 训练
def train_batch(X, y, device_params, devices, lr):
X_shards, y_shards = split_batch(X, y, devices) # 将小批量X和y均匀分布到各个gpu上
# 从小批量和共享的设备参数里拿到每个gpu上的小批量X,y和参数,
# 再把得到的(X_shard, device_W)放到LeNet里面,和y_shard求损失并求和
# ls是list,每一gpu上的损失
ls = [loss(lenet(X_shard, device_W),y_shard).sum()
for X_shard, y_shard, device_W in zip(
X_shards, y_shards, device_params)]
for l in ls:
l.backward() # 对每个gpu算梯度
with torch.no_grad():
for i in range(len(device_params[0])): # 这步结果是将每个设备参数的梯度存储的都是整体小批量的梯度
allreduce([device_params[c][i].grad
for c in range(len(devices))])
# 对每个梯度做同步更新优化
for param in device_params:
d2l.sgd(param, lr, X.shape[0])
定义训练函数
def train(num_gpus, batch_size, lr):
train_iter, test_iter = d2l.load_data_fashion_mnist(batch_size)
devices = [d2l.try_gpu(i) for i in range(num_gpus)]
# 主要区别在于要将参数复制到每个gpu上
device_params = [get_params(params, d) for d in devices] # 在训练之前将所有参数复制到每个gpu上
num_epochs = 10
animator = d2l.Animator('epoch', 'test acc', xlim=[1, num_epochs])
timer = d2l.Timer()
for epoch in range(num_epochs):
timer.start()
for X, y in train_iter:
train_batch(X, y, device_params, devices, lr)
torch.cuda.synchronize() # 同步一次
timer.stop()
animator.add(epoch + 1, (d2l.evaluate_accuracy_gpu(
lambda x: lenet(x, device_params[0]), test_iter, devices[0]),))
print(f'test acc: {animator.Y[0][-1]:.2f}, {timer.avg():.1f} sec/epoch '
f'on {str(devices)}')
# 在单个GPU上运行
train(num_gpus=1, batch_size=256, lr=0.2)
# 增加为2个GPU
train(num_gpus=2, batch_size=256, lr=0.2)
理论上不会改变训练结果,也有可能分布式也不会变快,
主要:gpu增加了,batch_size没变,会导致gpu性能不能完全发挥,如果只是单纯增加batch_size,可能会导致acc降低,
所以一般可以将批量和lr一起提高一点
2. 简洁实现
import torch
from torch import nn
from d2l import torch as d2l
# 使用一个稍微大一点的模型
def resnet18(num_classes, in_channels=1):
"""稍加修改的 ResNet-18 模型。"""
def resnet_block(in_channels, out_channels, num_residuals,
first_block=False):
blk = []
for i in range(num_residuals):
if i == 0 and not first_block:
blk.append(
d2l.Residual(in_channels, out_channels, use_1x1conv=True,
strides=2))
else:
blk.append(d2l.Residual(out_channels, out_channels))
return nn.Sequential(*blk)
net = nn.Sequential(
nn.Conv2d(in_channels, 64, kernel_size=3, stride=1, padding=1),
nn.BatchNorm2d(64), nn.ReLU())
net.add_module("resnet_block1", resnet_block(64, 64, 2, first_block=True))
net.add_module("resnet_block2", resnet_block(64, 128, 2))
net.add_module("resnet_block3", resnet_block(128, 256, 2))
net.add_module("resnet_block4", resnet_block(256, 512, 2))
net.add_module("global_avg_pool", nn.AdaptiveAvgPool2d((1, 1)))
net.add_module("fc",
nn.Sequential(nn.Flatten(), nn.Linear(512, num_classes)))
return net
net = resnet18(10)
devices = d2l.try_all_gpus()
训练-主要就是DataParallel的使用
DataParallel接口隐含的实现了get_params/ allreduce/split_batch等函数的功能
def train(net, num_gpus, batch_size, lr):
train_iter, test_iter = d2l.load_data_fashion_mnist(batch_size)
devices = [d2l.try_gpu(i) for i in range(num_gpus)]
def init_weights(m):
if type(m) in [nn.Linear, nn.Conv2d]:
nn.init.normal_(m.weight, std=0.01)
net.apply(init_weights)
# .DataParallel接口隐含的实现了get_params/ allreduce/split_batch等函数的功能
net = nn.DataParallel(net, device_ids=devices) # 把net复制到每个gpu上
trainer = torch.optim.SGD(net.parameters(), lr)
loss = nn.CrossEntropyLoss()
timer, num_epochs = d2l.Timer(), 10
animator = d2l.Animator('epoch', 'test acc', xlim=[1, num_epochs])
for epoch in range(num_epochs):
net.train()
timer.start()
for X, y in train_iter:
trainer.zero_grad()
X, y = X.to(devices[0]), y.to(devices[0]) # 和前面实现一样,要先将数据都复制到同一块gpu上
l = loss(net(X), y)
l.backward()
trainer.step()
timer.stop()
animator.add(epoch + 1, (d2l.evaluate_accuracy_gpu(net, test_iter),))
print(f'test acc: {animator.Y[0][-1]:.2f}, {timer.avg():.1f} sec/epoch '
f'on {str(devices)}')
# 在单个GPU上训练网络
train(net, num_gpus=1, batch_size=256, lr=0.1)
# 使用 2 个 GPU 进行训练
train(net, num_gpus=2, batch_size=512, lr=0.2)
可以通过调整批量和学习率来提升训练速度