- 🍨 本文为🔗365天深度学习训练营 中的学习记录博客
- 🍦 参考文章：Pytorch实战 | 第P6周：好莱坞明星识别
- 🍖 原作者：K同学啊 | 接辅导、项目定制
- 🚀 文章来源：K同学的学习圈子
本次采用的数据集为K同学提供的好莱坞明星数据集。
本人电脑配置
Python 3.8.0
Pytorch 1.8.1
torchvision 1.8.1+ cuda10.2
前期准备
1. 设置GPU/CPU
本次是在gpu上对网络进行训练和测试，先识别设备，判断设备类型。
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
print(device)
2. 导入数据
下载数据到主目录的文件夹的6-data文件夹里，该文件夹下共17个文件夹，1800个文件，读取文件后，对图片进行预处理，再进行训练集和测试集的划分，比例为4：1。
data_dir = './6-data/'
data_dir = pathlib.Path(data_dir)
data_paths = list(data_dir.glob('*'))
classNames = [str(path).split("\\")[1] for path in data_paths]
print(classNames)
data_transforms = transforms.Compose([
transforms.Resize([224,224]),
transforms.ToTensor(),
transforms.Normalize(
mean = [0.485,0.456,0.406],std = [0.229,0.224,0.225])
])
total_data = datasets.ImageFolder("./6-data/",transform=data_transforms)
print(total_data)
# divide dataset
train_size = int(0.8*len(total_data))
test_size = len(total_data) - train_size
train_dataset, test_dataset = torch.utils.data.random_split(total_data, [train_size,test_size])
print(len(train_dataset))
print(len(test_dataset))
VGG16网络模型
1. 搭建模型
VGG16网络是常用的分类网络结构，其网络结构如下图所示，在这里我们调用了官方给的预训练模型，把最后一层分类器修改，只训练最后一层参数。
from torchvision.models import vgg16
device = "cuda" if torch.cuda.is_available() else "cpu"
model = vgg16(pretrained = True).to(device)
for param in model.parameters():
# 冻结模型参数，只训练最后一层参数
param.requires_grad = False
model.classifier._modules['6'] = nn.Linear(4096, len(classNames))
model.to(device)
3. 编写训练函数
设置损失函数，这里采用的交叉熵损失函数，设置优化器为SGD优化，添加了学习率衰减策略，每隔4epoch，学习率会变为上一个值的92%。
# 训练循环
learn_rate = 1e-4 # 初始学习率
lambda1 = lambda epoch: 0.92 ** (epoch // 4)
loss_fn = nn.CrossEntropyLoss()
optimizer = torch.optim.SGD(model.parameters(), lr=learn_rate)
scheduler = torch.optim.lr_scheduler.LambdaLR(optimizer, lr_lambda=lambda1) #选定调整方法
# train model
def train(dataloader, model, loss_fn, optimizer):
size = len(dataloader.dataset)
num_batches = len(dataloader)
train_loss, train_acc = 0, 0
for X, y in dataloader:
X, y = X.to(device), y.to(device)
pred = model(X)
loss = loss_fn(pred, y)
optimizer.zero_grad()
loss.backward()
optimizer.step()
train_acc += (pred.argmax(1)==y).type(torch.float).sum().item()
train_loss += loss.item()
train_acc /= size
train_loss /= num_batches
return train_acc, train_loss
4. 编写测试函数
当不进行训练时，停止梯度更新，节省计算内存消耗。
# Test
def test(dataloader, model, loss_fn):
size = len(dataloader.dataset)
num_batches = len(dataloader)
test_loss, test_acc = 0, 0
with torch.no_grad():
for imgs, target in dataloader:
imgs, target = imgs.to(device), target.to(device)
target_pred = model(imgs)
loss = loss_fn(target_pred, target)
test_loss += loss.item()
test_acc += (target_pred.argmax(1)==target).type(torch.float).sum().item()
test_acc /= size
test_loss /= num_batches
return test_acc, test_loss
5. 主函数
设置迭代epoch次数，这里设定为40，并记录训练误差、精度，测试误差、精度。保存下最优的模型。
import copy
loss_fn = nn.CrossEntropyLoss()
epochs = 40
train_loss = []
train_acc = []
test_loss = []
test_acc = []
best_acc = 0
for epoch in range(epochs):
# 更新学习率，使用自定义学习率时使用
# adjust_learning_rate(optimizer, epoch, learn_rate)
model.train()
epoch_train_acc, epoch_train_loss = train(train_dl, model, loss_fn, optimizer)
scheduler.step()
model.eval()
epoch_test_acc, epoch_test_loss = test(test_dl, model, loss_fn)
if epoch_test_acc > best_acc:
best_acc = epoch_test_acc
best_model = copy.deepcopy(model)
train_acc.append(epoch_train_acc)
train_loss.append(epoch_train_loss)
test_acc.append(epoch_test_acc)
test_loss.append(epoch_test_loss)
lr = optimizer.state_dict()['param_groups'][0]['lr']
template = ('Epoch:{:2d}, Train_acc:{:.1f}%, Train_loss{:.3f}, Test_acc:{:.1f}%, Test_loss{:.3f}, Lr:{:.2E}')
print(template.format(epoch+1, epoch_train_acc*100, epoch_train_loss, epoch_test_acc*100, epoch_test_loss, lr))
PATH = './best_model_p6.pth'
torch.save(model.state_dict(),PATH)
结果总结
（1）采用原始网络进行训练得到的训练结果如下。可以看到训练精度大概再百分之二十多，比较低。同时注意到在epoch40以内，训练精度、测试精度、训练误差、测试误差都没有达到稳定值， 但有收敛趋势。可以稍微增大一下学习率。
（2）改进结构，使得精度达到60%以上。为了提高精度，我把分类器中的参数都进行了学习训练，并添加droupout层防止过拟合。首先看一下网络代码。
from torchvision import models
class VGGnet(nn.Module):
def __init__(self,feature_extract=True,num_classes=5):
super(VGGnet, self).__init__()
model = models.vgg16(pretrained=True)
self.features = model.features
set_parameter_requires_grad(self.features, feature_extract)#固定特征提取层参数
self.avgpool=model.avgpool
self.classifier = nn.Sequential(
nn.Linear(512 * 7 * 7, 512), #512 * 7 * 7不能改变 ，由VGG16网络决定的，第二个参数为神经元个数可以微调
nn.ReLU(True),
nn.Dropout(),
nn.Linear(512, 128),
nn.ReLU(True),
nn.Dropout(),
nn.Linear(128, len(classNames)),
)
def forward(self, x):
x = self.features(x)
x = self.avgpool(x)
x = x.view(x.size(0), 512*7*7)
out=self.classifier(x)
return out
def set_parameter_requires_grad(model, feature_extracting):
if feature_extracting:
for param in model.parameters():
param.requires_grad = False
同时对学习率的初始值和衰减率进行调整，初始值设定为1e-2，学习率=0.999*原来值
learn_rate = 1e-2
lambda1 = lambda epoch:0.999**(epoch//4)
optimizer = torch.optim.SGD(model.parameters(), lr = learn_rate)
得到的训练结果如下所示。可以看到相比于上面结果，精度有了大幅提高，最终在训练集上精度为64.4%。
每个epoch的精度和损失具体值如下。
Epoch: 1, Train_acc:13.6%, Train_loss2.686, Test_acc:19.7%, Test_loss2.448, Lr:1.00E-02
Epoch: 2, Train_acc:22.0%, Train_loss2.387, Test_acc:25.6%, Test_loss2.153, Lr:1.00E-02
Epoch: 3, Train_acc:27.8%, Train_loss2.114, Test_acc:37.2%, Test_loss1.965, Lr:1.00E-02
Epoch: 4, Train_acc:36.1%, Train_loss1.904, Test_acc:38.3%, Test_loss1.820, Lr:9.99E-03
Epoch: 5, Train_acc:43.4%, Train_loss1.702, Test_acc:45.0%, Test_loss1.701, Lr:9.99E-03
Epoch: 6, Train_acc:52.5%, Train_loss1.478, Test_acc:48.9%, Test_loss1.541, Lr:9.99E-03
Epoch: 7, Train_acc:57.7%, Train_loss1.312, Test_acc:50.3%, Test_loss1.484, Lr:9.99E-03
Epoch: 8, Train_acc:62.7%, Train_loss1.165, Test_acc:50.8%, Test_loss1.429, Lr:9.98E-03
Epoch: 9, Train_acc:67.6%, Train_loss1.007, Test_acc:53.1%, Test_loss1.321, Lr:9.98E-03
Epoch:10, Train_acc:73.1%, Train_loss0.896, Test_acc:58.6%, Test_loss1.294, Lr:9.98E-03
Epoch:11, Train_acc:75.9%, Train_loss0.782, Test_acc:58.1%, Test_loss1.228, Lr:9.98E-03
Epoch:12, Train_acc:80.9%, Train_loss0.668, Test_acc:57.8%, Test_loss1.199, Lr:9.97E-03
Epoch:13, Train_acc:83.7%, Train_loss0.563, Test_acc:60.8%, Test_loss1.168, Lr:9.97E-03
Epoch:14, Train_acc:86.2%, Train_loss0.512, Test_acc:58.3%, Test_loss1.174, Lr:9.97E-03
Epoch:15, Train_acc:88.2%, Train_loss0.451, Test_acc:63.6%, Test_loss1.106, Lr:9.97E-03
Epoch:16, Train_acc:90.9%, Train_loss0.377, Test_acc:64.2%, Test_loss1.159, Lr:9.96E-03
Epoch:17, Train_acc:91.8%, Train_loss0.331, Test_acc:61.9%, Test_loss1.169, Lr:9.96E-03
Epoch:18, Train_acc:92.4%, Train_loss0.306, Test_acc:62.8%, Test_loss1.170, Lr:9.96E-03
Epoch:19, Train_acc:94.6%, Train_loss0.248, Test_acc:61.4%, Test_loss1.197, Lr:9.96E-03
Epoch:20, Train_acc:94.6%, Train_loss0.241, Test_acc:64.2%, Test_loss1.190, Lr:9.95E-03
Epoch:21, Train_acc:96.0%, Train_loss0.195, Test_acc:63.9%, Test_loss1.153, Lr:9.95E-03
Epoch:22, Train_acc:96.0%, Train_loss0.194, Test_acc:64.2%, Test_loss1.147, Lr:9.95E-03
Epoch:23, Train_acc:96.4%, Train_loss0.166, Test_acc:64.7%, Test_loss1.160, Lr:9.95E-03
Epoch:24, Train_acc:97.6%, Train_loss0.140, Test_acc:65.0%, Test_loss1.109, Lr:9.94E-03
Epoch:25, Train_acc:97.1%, Train_loss0.131, Test_acc:63.9%, Test_loss1.178, Lr:9.94E-03
Epoch:26, Train_acc:97.1%, Train_loss0.135, Test_acc:64.2%, Test_loss1.223, Lr:9.94E-03
Epoch:27, Train_acc:98.4%, Train_loss0.104, Test_acc:64.7%, Test_loss1.163, Lr:9.94E-03
Epoch:28, Train_acc:98.3%, Train_loss0.098, Test_acc:63.3%, Test_loss1.190, Lr:9.93E-03
Epoch:29, Train_acc:98.3%, Train_loss0.092, Test_acc:65.6%, Test_loss1.148, Lr:9.93E-03
Epoch:30, Train_acc:99.0%, Train_loss0.077, Test_acc:66.9%, Test_loss1.182, Lr:9.93E-03
Epoch:31, Train_acc:98.1%, Train_loss0.090, Test_acc:65.6%, Test_loss1.141, Lr:9.93E-03
Epoch:32, Train_acc:99.2%, Train_loss0.068, Test_acc:63.3%, Test_loss1.246, Lr:9.92E-03
Epoch:33, Train_acc:99.2%, Train_loss0.064, Test_acc:64.7%, Test_loss1.174, Lr:9.92E-03
Epoch:34, Train_acc:98.9%, Train_loss0.067, Test_acc:64.2%, Test_loss1.256, Lr:9.92E-03
Epoch:35, Train_acc:99.2%, Train_loss0.056, Test_acc:66.1%, Test_loss1.222, Lr:9.92E-03
Epoch:36, Train_acc:98.8%, Train_loss0.065, Test_acc:64.4%, Test_loss1.284, Lr:9.91E-03
Epoch:37, Train_acc:99.3%, Train_loss0.058, Test_acc:64.2%, Test_loss1.288, Lr:9.91E-03
Epoch:38, Train_acc:99.3%, Train_loss0.053, Test_acc:64.2%, Test_loss1.293, Lr:9.91E-03
Epoch:39, Train_acc:99.2%, Train_loss0.053, Test_acc:64.4%, Test_loss1.275, Lr:9.91E-03
Epoch:40, Train_acc:99.5%, Train_loss0.047, Test_acc:64.4%, Test_loss1.207, Lr:9.90E-03
预测结果为：Megan Fox
0.6694444444444444
0.047233612628446685
（3）尝试把学习率初始值设置更大一点，但发现训练时极易出现过拟合的问题。随着epoch的增加，训练误差在逐渐减小，而测试误差在逐渐增大。可见学习率的值也不能设置过大。
Epoch: 1, Train_acc:14.2%, Train_loss2.613, Test_acc:21.4%, Test_loss2.246, Lr:2.00E-02
Epoch: 2, Train_acc:26.2%, Train_loss2.167, Test_acc:27.2%, Test_loss1.961, Lr:2.00E-02
Epoch: 3, Train_acc:34.5%, Train_loss1.881, Test_acc:35.8%, Test_loss1.831, Lr:2.00E-02
Epoch: 4, Train_acc:45.0%, Train_loss1.607, Test_acc:44.2%, Test_loss1.602, Lr:2.00E-02
Epoch: 5, Train_acc:52.4%, Train_loss1.373, Test_acc:44.7%, Test_loss1.453, Lr:2.00E-02
Epoch: 6, Train_acc:61.2%, Train_loss1.131, Test_acc:50.6%, Test_loss1.417, Lr:2.00E-02
Epoch: 7, Train_acc:68.0%, Train_loss0.941, Test_acc:48.6%, Test_loss1.456, Lr:2.00E-02
Epoch: 8, Train_acc:73.5%, Train_loss0.834, Test_acc:52.2%, Test_loss1.347, Lr:2.00E-02
Epoch: 9, Train_acc:78.5%, Train_loss0.676, Test_acc:52.2%, Test_loss1.363, Lr:2.00E-02
Epoch:10, Train_acc:82.0%, Train_loss0.574, Test_acc:54.4%, Test_loss1.363, Lr:2.00E-02
Epoch:11, Train_acc:86.4%, Train_loss0.480, Test_acc:56.7%, Test_loss1.346, Lr:2.00E-02
Epoch:12, Train_acc:88.1%, Train_loss0.396, Test_acc:53.3%, Test_loss1.447, Lr:1.99E-02
Epoch:13, Train_acc:91.0%, Train_loss0.341, Test_acc:56.4%, Test_loss1.419, Lr:1.99E-02
Epoch:14, Train_acc:91.7%, Train_loss0.281, Test_acc:59.2%, Test_loss1.473, Lr:1.99E-02
Epoch:15, Train_acc:92.6%, Train_loss0.242, Test_acc:55.3%, Test_loss1.421, Lr:1.99E-02
Epoch:16, Train_acc:94.8%, Train_loss0.209, Test_acc:57.8%, Test_loss1.450, Lr:1.99E-02
Epoch:17, Train_acc:94.4%, Train_loss0.193, Test_acc:58.6%, Test_loss1.402, Lr:1.99E-02
Epoch:18, Train_acc:96.0%, Train_loss0.161, Test_acc:54.4%, Test_loss1.541, Lr:1.99E-02
Epoch:19, Train_acc:95.9%, Train_loss0.146, Test_acc:57.5%, Test_loss1.470, Lr:1.99E-02
Epoch:20, Train_acc:96.9%, Train_loss0.123, Test_acc:54.4%, Test_loss1.494, Lr:1.99E-02
Epoch:21, Train_acc:97.4%, Train_loss0.098, Test_acc:57.5%, Test_loss1.551, Lr:1.99E-02
Epoch:22, Train_acc:98.1%, Train_loss0.095, Test_acc:57.5%, Test_loss1.580, Lr:1.99E-02
Epoch:23, Train_acc:97.9%, Train_loss0.077, Test_acc:56.1%, Test_loss1.594, Lr:1.99E-02
扫一扫
3077
被折叠的 条评论
为什么被折叠?
到【灌水乐园】发言
