為每個(gè)新模型從頭開(kāi)始實(shí)施并行性并不好玩。此外，優(yōu)化同步工具以獲得高性能有很大的好處。在下文中，我們將展示如何使用深度學(xué)習(xí)框架的高級(jí) API 來(lái)執(zhí)行此操作。數(shù)學(xué)和算法與第 13.5 節(jié)中的相同。毫不奇怪，您至少需要兩個(gè) GPU 才能運(yùn)行本節(jié)的代碼。

import torch
from torch import nn
from d2l import torch as d2l

from mxnet import autograd, gluon, init, np, npx
from mxnet.gluon import nn
from d2l import mxnet as d2l

npx.set_np()

13.6.1。玩具網(wǎng)絡(luò)

讓我們使用一個(gè)比13.5 節(jié)中的 LeNet 更有意義的網(wǎng)絡(luò) ，它仍然足夠容易和快速訓(xùn)練。我們選擇了一個(gè) ResNet-18 變體（He et al. , 2016）。由于輸入圖像很小，我們對(duì)其進(jìn)行了輕微修改。特別地，與第 8.6 節(jié)的不同之處在于，我們?cè)陂_(kāi)始時(shí)使用了更小的卷積核、步長(zhǎng)和填充。此外，我們刪除了最大池化層。

#@save
def resnet18(num_classes, in_channels=1):
  """A slightly modified ResNet-18 model."""
  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(out_channels, use_1x1conv=True,
                    strides=2))
      else:
        blk.append(d2l.Residual(out_channels))
    return nn.Sequential(*blk)

  # This model uses a smaller convolution kernel, stride, and padding and
  # removes the max-pooling layer
  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

#@save
def resnet18(num_classes):
  """A slightly modified ResNet-18 model."""
  def resnet_block(num_channels, num_residuals, first_block=False):
    blk = nn.Sequential()
    for i in range(num_residuals):
      if i == 0 and not first_block:
        blk.add(d2l.Residual(
          num_channels, use_1x1conv=True, strides=2))
      else:
        blk.add(d2l.Residual(num_channels))
    return blk

  net = nn.Sequential()
  # This model uses a smaller convolution kernel, stride, and padding and
  # removes the max-pooling layer
  net.add(nn.Conv2D(64, kernel_size=3, strides=1, padding=1),
      nn.BatchNorm(), nn.Activation('relu'))
  net.add(resnet_block(64, 2, first_block=True),
      resnet_block(128, 2),
      resnet_block(256, 2),
      resnet_block(512, 2))
  net.add(nn.GlobalAvgPool2D(), nn.Dense(num_classes))
  return net

13.6.2。網(wǎng)絡(luò)初始化

我們將在訓(xùn)練循環(huán)內(nèi)初始化網(wǎng)絡(luò)。有關(guān)初始化方法的復(fù)習(xí)，請(qǐng)參閱第 5.4 節(jié)。

net = resnet18(10)
# Get a list of GPUs
devices = d2l.try_all_gpus()
# We will initialize the network inside the training loop

The initialize function allows us to initialize parameters on a device of our choice. For a refresher on initialization methods see Section 5.4. What is particularly convenient is that it also allows us to initialize the network on multiple devices simultaneously. Let’s try how this works in practice.

net = resnet18(10)
# Get a list of GPUs
devices = d2l.try_all_gpus()
# Initialize all the parameters of the network
net.initialize(init=init.Normal(sigma=0.01), ctx=devices)

Using the split_and_load function introduced in Section 13.5 we can divide a minibatch of data and copy portions to the list of devices provided by the devices variable. The network instance automatically uses the appropriate GPU to compute the value of the forward propagation. Here we generate 4 observations and split them over the GPUs.

x = np.random.uniform(size=(4, 1, 28, 28))
x_shards = gluon.utils.split_and_load(x, devices)
net(x_shards[0]), net(x_shards[1])

[08:00:43] src/operator/nn/./cudnn/./cudnn_algoreg-inl.h:97: Running performance tests to find the best convolution algorithm, this can take a while... (set the environment variable MXNET_CUDNN_AUTOTUNE_DEFAULT to 0 to disable)

(array([[ 2.2610207e-06, 2.2045981e-06, -5.4046786e-06, 1.2869955e-06,
     5.1373163e-06, -3.8297967e-06, 1.4339059e-07, 5.4683451e-06,
     -2.8279192e-06, -3.9651104e-06],
    [ 2.0698672e-06, 2.0084667e-06, -5.6382510e-06, 1.0498458e-06,
     5.5506434e-06, -4.1065491e-06, 6.0830087e-07, 5.4521784e-06,
     -3.7365021e-06, -4.1891640e-06]], ctx=gpu(0)),
 array([[ 2.4629783e-06, 2.6015525e-06, -5.4362617e-06, 1.2938218e-06,
     5.6387889e-06, -4.1360108e-06, 3.5758853e-07, 5.5125256e-06,
     -3.1957325e-06, -4.2976326e-06],
    [ 1.9431673e-06, 2.2600434e-06, -5.2698201e-06, 1.4807417e-06,
     5.4830934e-06, -3.9678889e-06, 7.5751018e-08, 5.6764356e-06,
     -3.2530229e-06, -4.0943951e-06]], ctx=gpu(1)))

Once data passes through the network, the corresponding parameters are initialized on the device the data passed through. This means that initialization happens on a per-device basis. Since we picked GPU 0 and GPU 1 for initialization, the network is initialized only there, and not on the CPU. In fact, the parameters do not even exist on the CPU. We can verify this by printing out the parameters and observing any errors that might arise.

weight = net[0].params.get('weight')

try:
  weight.data()
except RuntimeError:
  print('not initialized on cpu')
weight.data(devices[0])[0], weight.data(devices[1])[0]

not initialized on cpu

(array([[[ 0.01382882, -0.01183044, 0.01417865],
     [-0.00319718, 0.00439528, 0.02562625],
     [-0.00835081, 0.01387452, -0.01035946]]], ctx=gpu(0)),
 array([[[ 0.01382882, -0.01183044, 0.01417865],
     [-0.00319718, 0.00439528, 0.02562625],
     [-0.00835081, 0.01387452, -0.01035946]]], ctx=gpu(1)))

Next, let’s replace the code to evaluate the accuracy by one that works in parallel across multiple devices. This serves as a replacement of the evaluate_accuracy_gpu function from Section 7.6. The main difference is that we split a minibatch before invoking the network. All else is essentially identical.

#@save
def evaluate_accuracy_gpus(net, data_iter, split_f=d2l.split_batch):
  """Compute the accuracy for a model on a dataset using multiple GPUs."""
  # Query the list of devices
  devices = list(net.collect_params().values())[0].list_ctx()
  # No. of correct predictions, no. of predictions
  metric = d2l.Accumulator(2)
  for features, labels in data_iter:
    X_shards, y_shards = split_f(features, labels, devices)
    # Run in parallel
    pred_shards = [net(X_shard) for X_shard in X_shards]
    metric.add(sum(float(d2l.accuracy(pred_shard, y_shard)) for
            pred_shard, y_shard in zip(
              pred_shards, y_shards)), labels.size)
  return metric[0] / metric[1]

13.6.3。訓(xùn)練

和以前一樣，訓(xùn)練代碼需要執(zhí)行幾個(gè)基本功能以實(shí)現(xiàn)高效并行：

需要在所有設(shè)備上初始化網(wǎng)絡(luò)參數(shù)。

在迭代數(shù)據(jù)集時(shí)，小批量將被劃分到所有設(shè)備上。

我們跨設(shè)備并行計(jì)算損失及其梯度。

梯度被聚合并且參數(shù)被相應(yīng)地更新。

最后，我們計(jì)算精度（再次并行）以報(bào)告網(wǎng)絡(luò)的最終性能。訓(xùn)練例程與前面章節(jié)中的實(shí)現(xiàn)非常相似，只是我們需要拆分和聚合數(shù)據(jù)。

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(module):
    if type(module) in [nn.Linear, nn.Conv2d]:
      nn.init.normal_(module.weight, std=0.01)
  net.apply(init_weights)
  # Set the model on multiple GPUs
  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'test acc: {animator.Y[0][-1]:.2f}, {timer.avg():.1f} sec/epoch '
     f'on {str(devices)}')

def train(num_gpus, batch_size, lr):
  train_iter, test_iter = d2l.load_data_fashion_mnist(batch_size)
  ctx = [d2l.try_gpu(i) for i in range(num_gpus)]
  net.initialize(init=init.Normal(sigma=0.01), ctx=ctx, force_reinit=True)
  trainer = gluon.Trainer(net.collect_params(), 'sgd',
              {'learning_rate': lr})
  loss = gluon.loss.SoftmaxCrossEntropyLoss()
  timer, num_epochs = d2l.Timer(), 10
  animator = d2l.Animator('epoch', 'test acc', xlim=[1, num_epochs])
  for epoch in range(num_epochs):
    timer.start()
    for features, labels in train_iter:
      X_shards, y_shards = d2l.split_batch(features, labels, ctx)
      with autograd.record():
        ls = [loss(net(X_shard), y_shard) for X_shard, y_shard
           in zip(X_shards, y_shards)]
      for l in ls:
        l.backward()
      trainer.step(batch_size)
    npx.waitall()
    timer.stop()
    animator.add(epoch + 1, (evaluate_accuracy_gpus(net, test_iter),))
  print(f'test acc: {animator.Y[0][-1]:.2f}, {timer.avg():.1f} sec/epoch '
     f'on {str(ctx)}')

讓我們看看這在實(shí)踐中是如何工作的。作為熱身，我們?cè)趩蝹€(gè) GPU 上訓(xùn)練網(wǎng)絡(luò)。

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

test acc: 0.90, 14.0 sec/epoch on [device(type='cuda', index=0)]

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

test acc: 0.93, 14.3 sec/epoch on [gpu(0)]

接下來(lái)我們使用 2 個(gè) GPU 進(jìn)行訓(xùn)練。與 13.5 節(jié)中評(píng)估的 LeNet 相比，ResNet-18 的模型要復(fù)雜得多。這就是并行化顯示其優(yōu)勢(shì)的地方。計(jì)算時(shí)間明顯大于同步參數(shù)的時(shí)間。這提高了可伸縮性，因?yàn)椴⑿谢拈_(kāi)銷(xiāo)不太相關(guān)。

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

test acc: 0.89, 8.8 sec/epoch on [device(type='cuda', index=0), device(type='cuda', index=1)]

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

test acc: 0.91, 14.2 sec/epoch on [gpu(0), gpu(1)]

13.6.4。概括

Gluon 通過(guò)提供上下文列表為跨多個(gè)設(shè)備的模型初始化提供原語(yǔ)。

在可以找到數(shù)據(jù)的設(shè)備上自動(dòng)評(píng)估數(shù)據(jù)。

在嘗試訪問(wèn)該設(shè)備上的參數(shù)之前，請(qǐng)注意初始化每個(gè)設(shè)備上的網(wǎng)絡(luò)。否則你會(huì)遇到錯(cuò)誤。

優(yōu)化算法自動(dòng)聚合多個(gè) GPU。

13.6.5。練習(xí)