寫在前面

限于財力不足，本機上只有一個 GPU 可供使用，因此這部分的代碼只能夠稍作了解，能夠使用的 GPU 也只有一個。

多 GPU 的數據并行：有幾張卡，對一個小批量數據，有幾張卡就分成幾塊，每個 GPU 分別計算梯度，然后加起來做并行。

從零開始實現

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

簡單網絡

# 初始化模型參數
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]# 定義模型
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# 交叉熵損失函數
loss = nn.CrossEntropyLoss(reduction='none')

向多個設備分發參數，并通過將模型參數復制到一個GPU：

def get_params(params, device): # 把一個參數復制到另外一個GPU上去new_params = [p.to(device) for p in params]for p in new_params:p.requires_grad_() #對每一個參數都需要計算梯度return new_paramsnew_params = get_params(params, d2l.try_gpu(0))
print('b1 權重:', new_params[1])
print('b1 梯度:', new_params[1].grad)

allreduce函數將所有向量相加，并將結果廣播給所有GPU

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])

將一個小批量數據均勻地分布在多個 GPU 上

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):"""將X和y拆分到多個設備上"""assert X.shape[0] == y.shape[0]return (nn.parallel.scatter(X, devices),nn.parallel.scatter(y, devices))

在一個小批量上實現多GPU訓練

def train_batch(X, y, device_params, devices, lr):X_shards, y_shards = split_batch(X, y, devices)# 在每個GPU上分別計算損失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:  # 反向傳播在每個GPU上分別執行l.backward()# 將每個GPU的所有梯度相加，并將其廣播到所有GPUwith torch.no_grad():for i in range(len(device_params[0])):allreduce([device_params[c][i].grad for c in range(len(devices))])# 在每個GPU上分別更新模型參數for param in device_params:d2l.sgd(param, lr, X.shape[0]) # 在這里，我們使用全尺寸的小批量

定義訓練模型：

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)]# 將模型參數復制到num_gpus個GPUdevice_params = [get_params(params, d) for d in devices]num_epochs = 10animator = 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:# 為單個小批量執行多GPU訓練train_batch(X, y, device_params, devices, lr)torch.cuda.synchronize()timer.stop()# 在GPU0上評估模型animator.add(epoch + 1, (d2l.evaluate_accuracy_gpu(lambda x: lenet(x, device_params[0]), test_iter, devices[0]),))print(f'測試精度：{animator.Y[0][-1]:.2f}，{timer.avg():.1f}秒/輪，'f'在{str(devices)}')

在單個 GPU 上運行：

增加為 2 個 GPU

并行后并沒有變快，可能有以下原因：

• Data 讀取比較慢
• GPU 增加了，但是 batch_size 沒有增加

多 GPU 的簡潔實現

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

簡單網絡

#@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)# 該模型使用了更小的卷積核、步長和填充，而且刪除了最大匯聚層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 netnet = resnet18(10)
# 獲取GPU列表
devices = d2l.try_all_gpus()
# 我們將在訓練代碼實現中初始化網絡

訓練

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)# 在多個GPU上設置模型net = nn.DataParallel(net, device_ids=devices)trainer = torch.optim.SGD(net.parameters(), lr)loss = nn.CrossEntropyLoss()timer, num_epochs = d2l.Timer(), 10animator = 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'測試精度：{animator.Y[0][-1]:.2f}，{timer.avg():.1f}秒/輪，'f'在{str(devices)}')

在單個 GPU 上訓練網絡

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

使用2個GPU進行訓練

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

QA 思考

Q1：驗證集準確率震蕩較大是哪個參數影響最大呢？
A1：lr

Q2：為什么batch_size調的比較小，比如8，精度會一直在0.1左右，一直不怎么變化
A2：因為batch_size調的比較小的時候，lr 不能太大。