《动手学》-单机多卡、分布式，batch_size和learning rate的设置说明

1.理论

4是绿线，5是橙线。这里key-value假设是内存。
注意这里batch_size是在每个

解释：可以显著地增大batch_size,同时也可相应地提高学习率。

2.代码

2.1从零自己实现

%matplotlib inline
import torch
from torch import nn
from torch.nn import functional as F
from d2l import torch as d2l

# 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')

数据同步

def get_params(params, device):
    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))
print('b1 weight:', new_params[1])
print('b1 grad:', new_params[1].grad)


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('before allreduce:\n', data[0], '\n', data[1])
allreduce(data)
print('after allreduce:\n', data[0], '\n', data[1])

数据分发

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)

#@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

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，因为放到某一个gpu上做，和放到各个gpu上做，效果都一样
    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

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GPU 0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'test acc: {animator.Y[0][-1]:.2f}, {timer.avg():.1f} sec/epoch '
          f'on {str(devices)}')


train(num_gpus=1, batch_size=256, lr=0.2)

train(num_gpus=2, batch_size=256, lr=0.2)

2.2 调用api简洁实现

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，而且删除了最大池化层。
    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

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)
    #不加这行代码可能会报错，先把模型放到gpu0上再做
    net.to(devices[0])
    # 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()
        #以下都会通过nn.DataParallel进行数据同步，数据分发操作
        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'test acc: {animator.Y[0][-1]:.2f}, {timer.avg():.1f} sec/epoch '
          f'on {str(devices)}')

train(net, num_gpus=1, batch_size=256, lr=0.1)

train(net, num_gpus=2, batch_size=512, lr=0.2)

3.QA

• Q：小批量分到多GPU计算后，模型结果怎么合到一起？
  梯度加起来后，模型只有一份，只在一个地方被更新。模型参数保证是一致的
• 每个GPU保存模型参数和梯度，因此，样本数会变少，常把batch_size调成gpu数的倍数
• 默认是将模型参数放到cpu的内存上，需要先挪到GPU
• 数据读入的时间是一定的，即使多GPU并行，这个时间也不会少，通信时间占比较大
• 数据多样性不够的时候，无法用大的batch_size的，因为里边会有很多相似的数据，算是无效数据
• learning rate可以稍微大一点，前期的抖动不影响后边的精度。
• Q：两个GPU训练时，最后的梯度是把两个GPU上的梯度相加吗？A：是的，因为mini-batch计算，就是多个样本的梯度求和的结果。
• Q：验证集上准确率震荡较大，主要是lr影响较大。

分布式

通过网络去传输数据，性能瓶颈在网络传输上。

【重点】

Q：batch_size越大，训练的有效性曲线是下降的？
A：极端情况，数据集就一个样本，复制了一万份。batch_size不管多大都是一样的，每个样本的梯度都是一样的，假设需要到某一个accuracy的话，如果batch_size为1，需要10个epoch才能到达，那么，即使batch_size为10，仍然需要10个epoch才能到达，而后者计算的数量是前者的10倍。所以，还是需要看数据集的多样性。