寫在前面
在本教程中,我們将學習:
- 通過
使用多GPU訓練模型.DataParallel
- 資料拟合.
使用多GPU
device = torch.device("cuda:0")
model.to(device) #傳回my_tensor的一個GPU上的備份, 而不是重寫覆寫了`my_tensor`,這種寫法是不正确的
mytensor = my_tensor.to(device) # 需要assign給一個新的tensor:mytensor,在GPU上用這個才合适
Pytorch預設隻使用一個GPU。代碼
model = nn.DataParallel(model)
将讓model在多個GPU上運作。
案例: 資料拟合
import torch
import torch.nn as nn
import torch.nn.functional as F
import torch.optim as optim
from torch.utils.data import Dataset, DataLoader
# Parameters and DataLoaders
# y(data_size) = X(data_size,2)xW(2,1) + b(标量)
# input_size = 2:代表設計矩陣 X 有2個特征
# output_size = 1:代表标簽y 隻有1個資料
input_size = 2
output_size = 1
batch_size = 200
data_size = 200
# 優先使用GPU
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
# Dummy DataSet
# 制作一個随機數資料集。你隻需要實作getitem函數
# @paras: length:資料集的長度(資料點的個數)
# n_features:資料的次元
class RandomDataset(Dataset):
def __init__(self, length,n_features):
self.len = length
self.weight = torch.tensor([[5.2],[9.6]])
self.bias = torch.tensor(3.4)
self.data = torch.randn(length, n_features)
self.targets = torch.matmul(self.data,self.weight) + self.bias
def __getitem__(self, index):
return self.data[index],self.targets[index]
def __len__(self):
return self.len
rand_loader = DataLoader(dataset=RandomDataset(data_size, 2),
batch_size=batch_size, shuffle=True)
# `DataParallel`可以用在任何模型上。
# 模型中的print語句将列印輸入tensor和輸出tensor的size.
# 注意batch rank0會列印什麼
class Model(nn.Module):
def __init__(self, input_size, output_size):
super(Model, self).__init__()
self.fc = nn.Linear(input_size, output_size)
def forward(self, input):
output = self.fc(input)
print("In Model: input size", input.size(),"output size", output.size())
return output
# 生成一個model執行個體:檢測是否有多個GPU
model = Model(input_size, output_size)
if torch.cuda.device_count() > 1:
print("Let's use", torch.cuda.device_count(), "GPUs!")
# dim = 0 [30, xxx] -> [10, ...], [10, ...], [10, ...] on 3 GPUs
model = nn.DataParallel(model)
model.to(device)
# 3 loss
# 4 optimizer
optimizer = optim.Adam(model.parameters(),lr=0.5)
def Train():
for epoch in range(200):
for (data,labels) in rand_loader:
# forward
input = data.to(device)
targets = labels.to(device)
output = model(input)
#loss = sum((output-targets)*(output-targets))/batch_size
loss = F.mse_loss(output,targets)
#if epoch%(20) == 0:
print("Outside: input size", input.size(),"output_size", output.size(),'loss:{}'.format(loss.item()))
# backward
optimizer.zero_grad()
loss.backward()
optimizer.step()
Train()
print(list(model.parameters()))
列印結果為(可見,PyTorch将資料集X均分到多個GPU上計算後,合并為輸出):
# If you have 2 GPUs, you will see:
# # on 2 GPUs
# Let's use 2 GPUs!
# In Model: input size torch.Size([100, 2]) output size torch.Size([100, 1])
# In Model: input size torch.Size([100, 2]) output size torch.Size([100, 1])
# Outside: input size torch.Size([200, 2]) output_size torch.Size([200, 1])
# In Model: input size torch.Size([100, 2]) output size torch.Size([100, 1])
# In Model: input size torch.Size([100, 2]) output size torch.Size([100, 1])
# Outside: input size torch.Size([200, 2]) output_size torch.Size([200, 1])
# ...