动手学习深度学习（14）深度学习硬件：计算性能

茴香豆 Lv5
  2022-11-04 15:31:36 2022-11-04 15:31:36 Created   2022-11-05 17:45:22 2022-11-05 17:45:22 Updated      

很好地理解算法和模型才可以捕获统计方面的问题，构建出具有出色性能的系统。同时，至少对底层硬件有一定的了解也是必不可少的。本节的内容可以作为理解某些算法为什么比其他算法更高效以及如何实现良好吞吐量的起点。

数据并行vs模型并行

  • 数据并行：将小批量分成n块，每个GPU拿到完整参数计算一块数据的梯度；通常性能更好
  • 模型并行：将模型分成n块，每个GPU拿到一块模型计算它的前向后后向结果；通常用于模型大到单GPU放不下

多GPU训练

1
2
3
4
import torch
from torch import nn
from torch.nn import functional as F
from d2l import torch as d2l

我们使用 6.6节中介绍的（稍加修改的）LeNet， 从零开始定义它，从而详细说明参数交换和同步。

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
# 初始化模型参数
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]

# 定义模型
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

# 交叉熵损失函数
loss = nn.CrossEntropyLoss(reduction='none')

向多个设备分发参数

1
2
3
4
5
6
7
8
9
10
11
12
def get_params(params, device):
    new_params = [p.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))
print('b1 权重:', new_params[1])
print('b1 梯度:', new_params[1].grad)
# output
b1 权重: tensor([0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.],
       device='cuda:0', requires_grad=True)
b1 梯度: None

allreduce函数将所有向量相加，并将结果广播给所有GPU

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
def allreduce(data):
    for i in range(1, len(data)):
        data[0][:] += data[i].to(data[0].device)
    for i in range(1, len(data)):
        data[i][:] = data[0].to(data[i].device)
data = [torch.ones((1, 2), device=d2l.try_gpu(i)) * (i + 1) for i in range(2)]
print('allreduce之前：\n', data[0], '\n', data[1])
allreduce(data)
print('allreduce之后：\n', data[0], '\n', data[1])
# output
allreduce之前：
 tensor([[1., 1.]], device='cuda:0')
 tensor([[2., 2.]], device='cuda:1')
allreduce之后：
 tensor([[3., 3.]], device='cuda:0')
 tensor([[3., 3.]], device='cuda:1')

我们需要一个简单的工具函数，将一个小批量数据均匀地分布在多个GPU上。 例如，有两个GPU时，我们希望每个GPU可以复制一半的数据。 因为深度学习框架的内置函数编写代码更方便、更简洁，所以在4×5矩阵上使用它进行尝试。

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
data = torch.arange(20).reshape(4, 5)
devices = [torch.device('cuda:0'), torch.device('cuda:1')]
split = nn.parallel.scatter(data, devices)
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'))

为了方便以后复用，我们定义了可以同时拆分数据和标签的split_batch函数。

1
2
3
4
5
6
#@save
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))

训练

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
def train_batch(X, y, device_params, devices, lr):
    X_shards, y_shards = split_batch(X, y, devices)
    # 在每个GPU上分别计算损失
    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:  # 反向传播在每个GPU上分别执行
        l.backward()
    # 将每个GPU的所有梯度相加，并将其广播到所有GPU
    with torch.no_grad():
        for i in range(len(device_params[0])):
            allreduce(
                [device_params[c][i].grad for c in range(len(devices))])
    # 在每个GPU上分别更新模型参数
    for param in device_params:
        d2l.sgd(param, lr, X.shape[0]) # 在这里，我们使用全尺寸的小批量
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
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)]
    # 将模型参数复制到num_gpus个GPU
    device_params = [get_params(params, d) for d in devices]
    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:
            # 为单个小批量执行多GPU训练
            train_batch(X, y, device_params, devices, lr)
            torch.cuda.synchronize()
        timer.stop()
        # 在GPU0上评估模型
        animator.add(epoch + 1, (d2l.evaluate_accuracy_gpu(
            lambda x: lenet(x, device_params[0]), test_iter, devices[0]),))
    print(f'测试精度：{animator.Y[0][-1]:.2f}，{timer.avg():.1f}秒/轮，'
          f'在{str(devices)}')
1
2
3
train(num_gpus=1, batch_size=256, lr=0.2)
# output
测试精度：0.80，2.7秒/轮，在[device(type='cuda', index=0)]
1
2
3
4
train(num_gpus=2, batch_size=256, lr=0.2)
# output
测试精度：0.84，2.8秒/轮，在[device(type='cuda', index=0), device(type='cuda', index=1)]

多GPU的简洁实现

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
import torch
from torch import nn
from d2l import torch as d2l

#@save
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)

    # 该模型使用了更小的卷积核、步长和填充，而且删除了最大汇聚层
    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)
# 获取GPU列表
devices = d2l.try_all_gpus()
# 我们将在训练代码实现中初始化网络

#训练
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)
    # 在多个GPU上设置模型
    net = nn.DataParallel(net, device_ids=devices)
    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])
            l = loss(net(X), y)
            l.backward()
            trainer.step()
        timer.stop()
        animator.add(epoch + 1, (d2l.evaluate_accuracy_gpu(net, test_iter),))
    print(f'测试精度：{animator.Y[0][-1]:.2f}，{timer.avg():.1f}秒/轮，'
          f'在{str(devices)}')
1
2
3
train(net, num_gpus=1, batch_size=256, lr=0.1)
# output
测试精度：0.92，13.7秒/轮，在[device(type='cuda', index=0)]
1
2
3
train(net, num_gpus=2, batch_size=512, lr=0.2)
# output
测试精度：0.89，8.4秒/轮，在[device(type='cuda', index=0), device(type='cuda', index=1)]
  • Title: 动手学习深度学习（14）深度学习硬件：计算性能
  • Author: 茴香豆
  • Created at : 2022-11-04 15:31:36
  • Updated at : 2022-11-05 17:45:22
  • Link: https://hxiangdou.github.io/2022/11/04/DL_14/
  • License: This work is licensed under CC BY-NC-SA 4.0.
Comments
On this page
动手学习深度学习（14）深度学习硬件：计算性能
  1. 数据并行vs模型并行
  2. 多GPU训练