先回到之前用的CNN网络进行手写数字识别的实验:
import torch
import torch.nn as nn
from torch.autograd import Variable
import torch.utils.data as Data
import torchvision #包括了一些数据库,图片的数据库也包含了
import matplotlib.pyplot as plt
#定义超参数
EPOCH = 1
BATCH_SIZE = 50
LR= 0.001
DOWNLOAD_MNIST = True
train_data = torchvision.datasets.MNIST(
root = r'D:\python\minist', #存储路径
train = True,
transform = torchvision.transforms.ToTensor(), #把下载的数据改成Tensor形式
#把(0-255)转换成(0-1)
download = DOWNLOAD_MNIST #如果没下载就确认下载
)
train_loader = Data.DataLoader(dataset = train_data,batch_size = BATCH_SIZE,shuffle = True)
#准备测试集
test_data = torchvision.datasets.MNIST(
root = r'D:\python\minist', #存储路径
train = False,#提取出来的不是training data,是test data
transform = torchvision.transforms.ToTensor(), #把下载的数据改成Tensor形式
#把(0-255)转换成(0-1)
download = False #如果已经下载了,就用False)
)
test_x = Variable(torch.unsqueeze(test_data.test_data,dim = 1),volatile = True).type(torch.FloatTensor)[:2000]/255
#把test_data换到0-1之间
test_y = test_data.test_labels[:2000]
#建立CNN神经网路
class CNN(nn.Module):
def __init__(self):
super(CNN,self).__init__()
self.conv1 = nn.Sequential(
nn.Conv2d(
in_channels = 1,#图像的高度
out_channels = 16,#filter的高度,提取出来16个特征放到后面去
kernel_size = 5,#filter为5*5,
stride = 1,#扫描两个相邻区域之间的步长
padding = 2 #在图片周围围上一圈0,使filter扫描的时候边缘不会出现不够的情况
#padding = (kenrel_size-stride )/2 = (5-1)/2 = 2
),#卷积层
nn.ReLU(),#激活函数
nn.MaxPool2d(kernel_size = 2),#池化层,筛选重要信息
)
self.conv2 = nn.Sequential(
nn.Conv2d(16,32,5,1,2),#卷积层
#前面输出16层,现在输入就是16层,输出就是32层
nn.ReLU(),#激活函数
nn.MaxPool2d(kernel_size = 2),#池化层,筛选重要信息
)
self.out = nn.Linear(32*7*7,10)
#输出是0-9十个类别的分类
#图片维度(1,28,28) -->conv2d --> (16,28,28) --> pa dding --> (16,14,14)
#-->(16,14,14) -->conv2d --> (32,14,14) --> padding --> (32,7,7)
#三维数据展平成2维数据
def forward(self,x):
x = self.conv1(x)
x = self.conv2(x)
x = x.view(x.size(0),-1) # 展平多维的卷积图成 (batch_size, 32 * 7 * 7)
output = self.out(x)
return output
cnn = CNN()
print(cnn) #打印结构
#优化器和loss
optimizer = torch.optim.Adam(cnn.parameters(),lr = LR)#优化器
loss_func = nn.CrossEntropyLoss()#计算损失函数
#训练过程
for epoch in range(EPOCH):
for step, (b_x, b_y) in enumerate(train_loader): # 分配 batch data, normalize x when iterate train_loader
b_x = Variable(b_x) #batch x
b_y = Variable(b_y) #batch y
output = cnn(b_x) # cnn output
loss = loss_func(output, b_y) # cross entropy loss
optimizer.zero_grad() # clear gradients for this training step
loss.backward() # backpropagation, compute gradients
optimizer.step() # apply gradients
#打印出来训练效果
if step % 50 == 0:
test_output = cnn(test_x)
pred_y = torch.max(test_output,1)[1].data.squeeze()
accuracy = float((pred_y == test_y).numpy().astype(int).sum()) / float(test_y.size(0))
#算括号里的是否等于,等于表示预测对了记一次,总共对的次数除以总数就是accuracy
print('Epoch: ',epoch,'| train loss: %.4f' % loss.data[0],'| test accuracy: %.2f' % accuracy)
#拿测试集前十个数据测试一下效果
test_output = cnn(test_x[:10])
pred_y = torch.max(test_output,1)[1].data.numpy().squeeze()
print(pred_y,'prediction number')
print(test_y[:10].numpy(),'real number')
为了比较出gpu和cpu运算的时间不同,在开头引入time模块
然后在训练的时候,开头末尾计时:
#训练过程
t_start = time.time()
for epoch in range(EPOCH):
for step, (b_x, b_y) in enumerate(train_loader): # 分配 batch data, normalize x when iterate train_loader
b_x = Variable(b_x) #batch x
b_y = Variable(b_y) #batch y
output = cnn(b_x) # cnn output
loss = loss_func(output, b_y) # cross entropy loss
optimizer.zero_grad() # clear gradients for this training step
loss.backward() # backpropagation, compute gradients
optimizer.step() # apply gradients
#打印出来训练效果
if step % 50 == 0:
test_output = cnn(test_x)
pred_y = torch.max(test_output,1)[1].data.squeeze()
accuracy = float((pred_y == test_y).numpy().astype(int).sum()) / float(test_y.size(0))
#算括号里的是否等于,等于表示预测对了记一次,总共对的次数除以总数就是accuracy
print('Epoch: ',epoch,'| train loss: %.4f' % loss.data[0],'| test accuracy: %.2f' % accuracy)
t_end = time.time()
最后通过print('time:',t_end-t_start)
算的时间:
现在使用gpu加速:
在cpu计算的基础上有几个需要改动的地方:
- 首先cnn网络必须移到gpu上去,移动方法很简单,直接在调用的cnn后.cuda():
cnn = CNN()
#print(cnn) #打印结构
#cnn模块也要移动到cuda上面去!!!!!!!!!!!
cnn.cuda()
- 测试数据test_y,test_x需要移动到cuda上:
test_x = Variable(torch.unsqueeze(test_data.test_data,dim = 1),volatile = True).type(torch.FloatTensor)[:2000].cuda()/255
#把test_data换到0-1之间
test_y = test_data.test_labels[:2000].cuda()
- 训练过程中训练数据b_x,b_y需要转移到cuda上去,直接调用cuda()就可以了:
#!!!!!!!!!!!!!!!!!!!!!!
b_x = Variable(b_x).cuda() #batch x #把原来的数据转移到gpu上面去
b_y = Variable(b_y).cuda() #batch y
- 在cuda上计算的accuracy和pred_y也需要改动:
但是需要注意的是,gpu上面不能进行numpy的操作,所以在gpu上进行数据计算的时候一定要注意数据类型是numpy的还是tensor的
#!!!!!!!!!!!!!!!!!!
pred_y = torch.max(test_output,1)[1].cuda().data.squeeze()
accuracy = torch.sum(pred_y == test_y).float()/float(test_y.size(0))
#算括号里的是否等于,等于表示预测对了记一次,总共对的次数除以总数就是accuracy
#.float()表示把int的tensor强制转换成float
#除号后面也是一个int型的数,float强制类型转换
- 最后打印的时候,用的numpy打印,所以应该提前把test_y变成cpu上的:
在gpu上处理时不能出现.numpy()
,如果是在要用numpy先用.cpu()
把数据转到cpu上面去
#!!!!!!!!!!!!!!!!!
#在打印之前要变回到cpu上
test_y = test_y.cpu()
print(pred_y,'prediction number')
print(test_y[:10].numpy(),'real number')
print('time:',t_end-t_start)
时间只用了7.5s,而在cpu上需要110.7s
gpu加速完整代码:
# -*- coding: utf-8 -*-
"""
Created on Sun Jul 19 10:58:07 2020
@author: lenovo
"""
# -*- coding: utf-8 -*-
"""
Created on Sat Jul 18 21:43:07 2020
@author: lenovo
"""
import torch
import torch.nn as nn
from torch.autograd import Variable
import torch.utils.data as Data
import torchvision #包括了一些数据库,图片的数据库也包含了
import matplotlib.pyplot as plt
import time
#定义超参数
EPOCH = 1
BATCH_SIZE = 50
LR= 0.001
DOWNLOAD_MNIST = True
train_data = torchvision.datasets.MNIST(
root = r'D:\python\minist', #存储路径
train = True,
transform = torchvision.transforms.ToTensor(), #把下载的数据改成Tensor形式
#把(0-255)转换成(0-1)
download = DOWNLOAD_MNIST #如果没下载就确认下载
)
train_loader = Data.DataLoader(dataset = train_data,batch_size = BATCH_SIZE,shuffle = True)
#准备测试集
test_data = torchvision.datasets.MNIST(
root = r'D:\python\minist', #存储路径
train = False,#提取出来的不是training data,是test data
transform = torchvision.transforms.ToTensor(), #把下载的数据改成Tensor形式
#把(0-255)转换成(0-1)
download = False #如果已经下载了,就用False)
)
#!!!!!!!!!!!!!!!!!!!!!!!1
test_x = Variable(torch.unsqueeze(test_data.test_data,dim = 1),volatile = True).type(torch.FloatTensor)[:2000].cuda()/255
#把test_data换到0-1之间
test_y = test_data.test_labels[:2000].cuda()
#建立CNN神经网路
class CNN(nn.Module):
def __init__(self):
super(CNN,self).__init__()
self.conv1 = nn.Sequential(
nn.Conv2d(
in_channels = 1,#图像的高度
out_channels = 16,#filter的高度,提取出来16个特征放到后面去
kernel_size = 5,#filter为5*5,
stride = 1,#扫描两个相邻区域之间的步长
padding = 2 #在图片周围围上一圈0,使filter扫描的时候边缘不会出现不够的情况
#padding = (kenrel_size-stride )/2 = (5-1)/2 = 2
),#卷积层
nn.ReLU(),#激活函数
nn.MaxPool2d(kernel_size = 2),#池化层,筛选重要信息
)
self.conv2 = nn.Sequential(
nn.Conv2d(16,32,5,1,2),#卷积层
#前面输出16层,现在输入就是16层,输出就是32层
nn.ReLU(),#激活函数
nn.MaxPool2d(kernel_size = 2),#池化层,筛选重要信息
)
self.out = nn.Linear(32*7*7,10)
#输出是0-9十个类别的分类
#图片维度(1,28,28) -->conv2d --> (16,28,28) --> pa dding --> (16,14,14)
#-->(16,14,14) -->conv2d --> (32,14,14) --> padding --> (32,7,7)
#三维数据展平成2维数据
def forward(self,x):
x = self.conv1(x)
x = self.conv2(x)
x = x.view(x.size(0),-1) # 展平多维的卷积图成 (batch_size, 32 * 7 * 7)
output = self.out(x)
return output
cnn = CNN()
print(cnn) #打印结构
#!!!!!!!!!!!!!!!!!!!!
cnn.cuda()
#优化器和loss
optimizer = torch.optim.Adam(cnn.parameters(),lr = LR)#优化器
loss_func = nn.CrossEntropyLoss()#计算损失函数
#训练过程
t_start = time.time()
for epoch in range(EPOCH):
for step, (b_x, b_y) in enumerate(train_loader): # 分配 batch data, normalize x when iterate train_loader
#!!!!!!!!!!!!!!
b_x = Variable(b_x).cuda() #batch x
b_y = Variable(b_y).cuda() #batch y
output = cnn(b_x) # cnn output
loss = loss_func(output, b_y) # cross entropy loss
optimizer.zero_grad() # clear gradients for this training step
loss.backward() # backpropagation, compute gradients
optimizer.step() # apply gradients
#打印出来训练效果
if step % 50 == 0:
test_output = cnn(test_x)
#!!!!!!!!!!!!!!!!!!
pred_y = torch.max(test_output,1)[1].cuda().data.squeeze()
accuracy = torch.sum(pred_y == test_y).float()/float(test_y.size(0))
#算括号里的是否等于,等于表示预测对了记一次,总共对的次数除以总数就是accuracy
#.float()表示把int的tensor强制转换成float
#除号后面也是一个int型的数,float强制类型转换
print('Epoch: ',epoch,'| train loss: %.4f' % loss.data[0],'| test accuracy: %.2f' % accuracy)
t_end = time.time()
#拿测试集前十个数据测试一下效果
test_output = cnn(test_x[:10])
pred_y = torch.max(test_output,1)[1].data.squeeze()
#!!!!!!!!!!!!!!!!!
#在打印之前要变回到cpu上
test_y = test_y.cpu()
print(pred_y,'prediction number')
print(test_y[:10].numpy(),'real number')
print('time:',t_end-t_start)