Ubuntu16.04 install cpu version of pytorch and test small demo

Foreword:

When you are familiar with installing Anaconda on ubuntu16.04 and using Anaconda to create a virtual environment corresponding to the python version , we start the next steps.

1. Install typroch

To the official website generate a corresponding typorch the pip or conda command

to open a terminal input:

source activate learn #切换到自己创建的learn环境，不会请看前言的两个博客
conda install pytorch torchvision cpuonly -c pytorch #安装pytorch（根据提示来即可）,若因某种原因出现段错误（核心已转储）的情况，重启虚拟机，再来一次即可。

Verify if there are errors after completion

python
import torch
import torchvision
#若没有保存则安装成功

2. Test the small demo:

# encoding: utf-8
import torch
import torch.nn as nn
import torch.nn.functional as F #加载nn中的功能函数
import torch.optim as optim #加载优化器有关包
import torch.utils.data as Data
from torchvision import datasets,transforms #加载计算机视觉有关包
from torch.autograd import Variable

BATCH_SIZE = 64

#加载torchvision包内内置的MNIST数据集 这里涉及到transform:将图片转化成torchtensor
train_dataset = datasets.MNIST(root='~/data/',train=True,transform=transforms.ToTensor(),download=True)
test_dataset = datasets.MNIST(root='~/data/',train=False,transform=transforms.ToTensor())

#加载小批次数据，即将MNIST数据集中的data分成每组batch_size的小块，shuffle指定是否随机读取
train_loader = Data.DataLoader(dataset=train_dataset,batch_size=BATCH_SIZE,shuffle=True)
test_loader = Data.DataLoader(dataset=test_dataset,batch_size=BATCH_SIZE,shuffle=False)

#定义网络模型亦即Net 这里定义一个简单的全连接层784->10
class Model(nn.Module):
    def __init__(self):
        super(Model,self).__init__()
        self.linear1 = nn.Linear(784,10)

    def forward(self,X):
        return F.relu(self.linear1(X))

model = Model() #实例化全连接层
loss = nn.CrossEntropyLoss() #损失函数选择，交叉熵函数
optimizer = optim.SGD(model.parameters(),lr = 0.1)
num_epochs = 5

#以下四个列表是为了可视化（暂未实现）
losses = []
acces = []
eval_losses = []
eval_acces = []

for echo in range(num_epochs):
    train_loss = 0   #定义训练损失
    train_acc = 0    #定义训练准确度
    model.train()    #将网络转化为训练模式
    for i,(X,label) in enumerate(train_loader):     #使用枚举函数遍历train_loader
        X = X.view(-1,784)       #X:[64,1,28,28] -> [64,784]将X向量展平
        X = Variable(X)          #包装tensor用于自动求梯度
        label = Variable(label)
        out = model(X)           #正向传播
        lossvalue = loss(out,label)         #求损失值
        optimizer.zero_grad()       #优化器梯度归零
        lossvalue.backward()    #反向转播，刷新梯度值
        optimizer.step()        #优化器运行一步，注意optimizer搜集的是model的参数

        #计算损失
        train_loss += float(lossvalue)
        #计算精确度
        _,pred = out.max(1)
        num_correct = (pred == label).sum()
        acc = float(num_correct) / X.shape[0]
        train_acc += acc

    losses.append(train_loss / len(train_loader))
    acces.append(train_acc / len(train_loader))
    print("echo:"+' ' +str(echo))
    print("lose:" + ' ' + str(train_loss / len(train_loader)))
    print("accuracy:" + ' '+str(train_acc / len(train_loader)))
    eval_loss = 0
    eval_acc = 0
    model.eval() #模型转化为评估模式
    for X,label in test_loader:
        X = X.view(-1,784)
        X = Variable(X)
        label = Variable(label)
        testout = model(X)
        testloss = loss(testout,label)
        eval_loss += float(testloss)

        _,pred = testout.max(1)
        num_correct = (pred == label).sum()
        acc = float(num_correct) / X.shape[0]
        eval_acc += acc

    eval_losses.append(eval_loss / len(test_loader))
    eval_acces.append(eval_acc / len(test_loader))
    print("testlose: " + str(eval_loss/len(test_loader)))
    print("testaccuracy:" + str(eval_acc/len(test_loader)) + '\n')

Pull test.py into the learn environment you created (you can put it anywhere, I just want to write less path) here

Open terminal input:

source activate learn #切换到自己创建的learn环境，不会请看前言的两个博客
python test.py

output: