本系列文章配套代码获取有以下两种途径：
-
通过百度网盘获取：
链接：https://pan.baidu.com/s/1XuxKa9_G00NznvSK0cr5qw?pwd=mnsj提取码：mnsj
-
前往GitHub获取：
https://github.com/returu/PyTorch【深度学习(PyTorch篇)】32.使用卷积神经网络完成图像分类任务（CIFAR-10数据集）
具体代码如下：
import torch
import torch.nn as nn
import torch.optim as optim
from torch.utils.data import DataLoader
from torchvision import datasets, transforms
from torchvision.utils import make_grid
import torch.nn.functional as F
import visdom
# 设置设备
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
# 定义模型
class Net(nn.Module):
def __init__(self):
super().__init__()
self.conv1 = nn.Conv2d(3, 6, 5)
self.bn1 = nn.BatchNorm2d(6)
self.pool = nn.MaxPool2d(2, 2)
self.conv2 = nn.Conv2d(6, 16, 5)
self.bn2 = nn.BatchNorm2d(16)
self.fc1 = nn.Linear(16 * 5 * 5, 120)
self.fc2 = nn.Linear(120, 84)
self.fc3 = nn.Linear(84, 10)
def forward(self, x):
x = self.pool(nn.functional.relu(self.bn1(self.conv1(x))))
x = self.pool(nn.functional.relu(self.bn2(self.conv2(x))))
x = x.view(-1, 16 * 5 * 5)
x = nn.functional.relu(self.fc1(x))
x = nn.functional.relu(self.fc2(x))
x = self.fc3(x)
return x
# 初始化模型和优化器
net = Net().to(device)
criterion = nn.CrossEntropyLoss()
optimizer = optim.SGD(net.parameters(), lr=0.001, momentum=0.9)
# 数据预处理
transform = transforms.Compose([transforms.ToTensor(), transforms.Normalize((0.4915,0.4823,0.4468),(0.2470,0.2435,0.2616))])
# 加载数据集
trainset = datasets.CIFAR10('./data/' , train=True , download=True,transform=transform)
train_loader = DataLoader(trainset , batch_size=64 , shuffle=True)
testset = datasets.CIFAR10('./data/' , train=False , download=True,transform=transform)
test_loader = DataLoader(testset , batch_size=64 , shuffle=True)
# 训练网络
for epoch in range(10): # 多次遍历整个数据集
running_loss = 0.0
correct = 0
total = 0
for i, data in enumerate(train_loader, 0):
# 获取输入数据
inputs, labels = data[0].to(device), data[1].to(device)
# 清零梯度缓存
optimizer.zero_grad()
# 前向传播，后向传播，优化
outputs = net(inputs)
loss = criterion(outputs, labels)
loss.backward()
optimizer.step()
# 输出统计信息
# 累计损失和准确率
running_loss += loss.item()
_, predicted = torch.max(outputs, 1)
total += labels.size(0)
correct += (predicted == labels).sum().item()
# 每200个mini-batches输出一次统计信息
if (i+1) % 200 == 0:
train_acc = 100 * correct / total
print(f'[{epoch + 1}, {i + 1}] loss: {(running_loss / 200):.3f}, train acc: {train_acc:.2f}')
running_loss = 0.0
correct = 0
total = 0
print('Finished Training')
# 测试网络
correct = 0
total = 0
loss_sum = 0.0
with torch.no_grad():
for data in test_loader:
images, labels = data[0].to(device), data[1].to(device)
outputs = net(images)
loss = criterion(outputs, labels) # 计算损失
loss_sum += loss.item() # 累加损失值
_, predicted = torch.max(outputs.data, 1)
total += labels.size(0)
correct += (predicted == labels).sum().item()
# 计算平均损失
average_loss = loss_sum / len(test_loader)
print(f'Accuracy of the network on the 10000 test images: {100 * correct / total}%')
print(f'Average loss on the 10000 test images: {average_loss}')
本次将在上述构建的简单的卷积神经网络基础上，分别使用TensorBoard和Visodm来可视化以下内容：
-
训练过程中的损失和准确率； -
各层权重参数的直方图； -
第一个卷积层的输出特征图。
TensorBoard：
# 设置设备
# ... 代码保持不变 ...
# 初始化TensorBoard的SummaryWriter
writer = SummaryWriter(log_dir = 'runs')
# 定义模型
# ... 代码保持不变 ...
# 初始化模型和优化器
# ... 代码保持不变 ...
# 数据预处理
# ... 代码保持不变 ...
# 加载数据集
# ... 代码保持不变 ...
# 训练网络
for epoch in range(10): # 多次遍历整个数据集
# ... 代码保持不变 ...
# 每200个mini-batches输出一次统计信息 + 更新一次Visdom
if (i+1) % 200 == 0:
train_acc = 100 * correct / total
print(f'[{epoch + 1}, {i + 1}] loss: {(running_loss / 200):.3f}, train acc: {train_acc:.2f}')
# 更新TensorBoard的损失和准确率
writer.add_scalar('Loss/train', running_loss / 200, epoch * len(train_loader) + i + 1)
writer.add_scalar('Accuracy/train', train_acc, epoch * len(train_loader) + i + 1)
running_loss = 0.0
correct = 0
total = 0
# 关闭TensorBoard的SummaryWriter
writer.close()
print('Finished Training')
# 各层权重参数直方图
for name, param in net.named_parameters():
if 'weight' in name:
writer.add_histogram(f'{name} Weights' , param.data.cpu().numpy().flatten())
# 测试网络
# ... 代码保持不变 ...
# 可视化第一个卷积层的输出特征图
dataiter = iter(train_loader)
images, labels = next(dataiter)
images = images.to(device)
features = net.conv1(images) # torch.Size([64, 6, 28, 28])
writer.add_images('Features' , features.cpu()[:,:1,:,:])
Visdom：
# 设置设备
# ... 代码保持不变 ...
# 初始化Visdom
vis = visdom.Visdom(env='TEST_Visdom')
# 定义模型
# ... 代码保持不变 ...
# 初始化模型和优化器
# ... 代码保持不变 ...
# 数据预处理
# ... 代码保持不变 ...
# 加载数据集
# ... 代码保持不变 ...
# 训练网络
for epoch in range(10): # 多次遍历整个数据集
# ... 代码保持不变 ...
# 每200个mini-batches输出一次统计信息 + 更新一次Visdom
if (i+1) % 200 == 0:
train_acc = 100 * correct / total
print(f'[{epoch + 1}, {i + 1}] loss: {(running_loss / 200):.3f}, train acc: {train_acc:.2f}')
# 更新Visdom的损失和准确率图表
vis.line(X=np.array([epoch * len(train_loader) + i + 1]),
Y=np.array([(running_loss / 200)]),
win='window_loss',
update='append',
opts=dict(title='Loss over time', xlabel='Iteration', ylabel='Loss'))
vis.line(X=np.array([epoch * len(train_loader) + i + 1]),
Y=np.array([train_acc]),
win='window_acc',
update='append',
opts=dict(title='Accuracy over time', xlabel='Iteration', ylabel='Accuracy'))
running_loss = 0.0
correct = 0
total = 0
print('Finished Training')
# 各层权重参数直方图
for name, param in net.named_parameters():
if 'weight' in name:
vis.histogram(param.data.cpu().numpy().flatten(), opts=dict(title=f'{name} Weights'))
# 测试网络
# ... 代码保持不变 ...
# 可视化第一个卷积层的输出特征图
dataiter = iter(train_loader)
images, labels = next(dataiter)
images = images.to(device)
features = net.conv1(images) # torch.Size([64, 6, 28, 28])
vis.images(features.cpu()[:,:1,:,:], nrow=8,opts=dict(title=f'Features'))
可视化结果如下所示：
更多内容可以前往官网查看：
https://pytorch.org/
本篇文章来源于微信公众号: 码农设计师
