transfer learning

Transfer learning 迁移学习

介绍

因为拥有足够大的数据集相对很少，因此通常在很大的数据集上对ConvNet进行预训练，然后将ConvNet用作初始化或固定特征提取器以完成感兴趣的任务。
三种常见的transfer learning的方法

（1）ConvNet作为固定特征提取器
删除最后一个完全连接层，然后将其余剩下的部分看作固定特征提取器

（2）微调ConvNet
通过在新的数据集上面重新训练分类器，可以实现
1.对所有层的微调
2.对部分更高层的微调（固定较早的层）

（3）预训练模型

pytorch教程示例（TRANSFER LEARNING FOR COMPUTER VISION TUTORIAL）

目的

训练一个模型对蚂蚁和蜜蜂进行分类数据集下载

数据集结构：

1. 图像预处理

因为教程中图片的训练集数量为120张图像，验证集有75张。这是一个比较小的数据集，因此需要图像预处理对数据集进行增强。
训练模型对象基于ImageNet,因此图片大小应该预处理为 224*224

data_transforms = {
‘train’: transforms.Compose([
#有augment的效果，将原始图片通过尺寸和纵横比随机变化后再调整为给定大小224
transforms.RandomResizedCrop(224),
transforms.RandomHorizontalFlip(), #以默认0.5的概率随机反转图像
transforms.ToTensor(), #归一化之前必须转成tensor
transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225]) #归一化
]),
#验证机没有augment
‘val’: transforms.Compose([
transforms.Resize(256),
transforms.CenterCrop(224),
transforms.ToTensor(),
transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
]),
}

关于transforms.Normalize(mean,std)

归一化的具体算式：image = (image - mean) / std
当mean(0.5,0.5,0.5)，std(0.5,0.5,0.5)则将图片在（-1，1）内归一化

2.定义Dataset，Dataloader

image_datasets = {
‘train’: datasets.ImageFolder(os.path.join(data_dir, ‘train’), data_transforms[‘train’])
‘val’: darasets.imageFolder(os.path.join(data_dir,‘val’),data_transforms[‘val’])
}
dataloaders = {
‘train’: torch.utils.data.DataLoader(image_datasets[‘train’], batch_size=4,
shuffle=True, num_workers=4)
‘val’: torch.utils.data.DataLoader(image_datasets[train], batch_size=4,
shuffle=True, num_workers=4)
}
dataset_sizes = {x: len(image_datasets[x]) for x in [‘train’, ‘val’]}
class_names = image_datasets[‘train’].classes
device = torch.device(“cuda:0” if torch.cuda.is_available() else “cpu”)

3.训练模型

编写函数train_model进行模型的训练

#伪代码
for epoch in range(n_epochs):
for data, targets in trainloader:
#每次反向传播之前需要将梯度设置为0
#因为pytorch会在随后的loss.backwar()中累计梯度
optimizer.zero_grad()
# Generate predictions
out = model(data)
# Calculate loss
loss = criterion(out, lables)
# Backpropagation
loss.backward()
# Update model parameters
optimizer.step()

具体代码：

def train_model(model, criterion, optimizer, scheduler, num_epochs=25):
    since = time.time()

    best_model_wts = copy.deepcopy(model.state_dict())
    best_acc = 0.0

    for epoch in range(num_epochs):
    	#记录每一轮输出
        print('Epoch {}/{}'.format(epoch, num_epochs - 1))
        print('-' * 10)

        # Each epoch has a training and validation phase
        for phase in ['train', 'val']:
            if phase == 'train':
                model.train()  # Set model to training mode
            else:
                model.eval()   # Set model to evaluate mode

            running_loss = 0.0
            running_corrects = 0

            # Iterate over data.
            for inputs, labels in dataloaders[phase]:
                inputs = inputs.to(device)
                labels = labels.to(device)

                #将梯度置0
                optimizer.zero_grad()

                # forward
                # track history if only in train
                with torch.set_grad_enabled(phase == 'train'):
                    outputs = model(inputs)
                    _, preds = torch.max(outputs, 1)
                    loss = criterion(outputs, labels)

                    #只是在train的阶段进行后向传播和优化器参数更新
                    if phase == 'train':
                        loss.backward()
                        optimizer.step()

                # statistics
                running_loss += loss.item() * inputs.size(0)
                running_corrects += torch.sum(preds == labels.data)
            if phase == 'train':
                scheduler.step()

            epoch_loss = running_loss / dataset_sizes[phase]
            epoch_acc = running_corrects.double() / dataset_sizes[phase]

            print('{} Loss: {:.4f} Acc: {:.4f}'.format(
                phase, epoch_loss, epoch_acc))

            # deep copy the model
            if phase == 'val' and epoch_acc > best_acc:
                best_acc = epoch_acc
                best_model_wts = copy.deepcopy(model.state_dict())

        print()

    time_elapsed = time.time() - since
    print('Training complete in {:.0f}m {:.0f}s'.format(
        time_elapsed // 60, time_elapsed % 60))
    print('Best val Acc: {:4f}'.format(best_acc))

    # load best model weights
    model.load_state_dict(best_model_wts)
    return model

4.微调convnet

本教程采用ResNet-18模型进行训练

#采用模型resnet18
model_ft = models.resnet18(pretrained=True)
#num_fitrs表示线性层的输入数量
num_ftrs = model_ft.fc.in_features
# 这里class设置为2,分为蜜蜂和蚂蚁两种
model_ft.fc = nn.Linear(num_ftrs, 2)

model_ft = model_ft.to(device)  #复制到gpu上

criterion = nn.CrossEntropyLoss()  #选择loss函数

# 选择优化器SGD
optimizer_ft = optim.SGD(model_ft.parameters(), lr=0.001, momentum=0.9)

#scheduler负责调节训练过程中的学习率
# 每7个epoch就衰减learning rate
exp_lr_scheduler = lr_scheduler.StepLR(optimizer_ft, step_size=7, gamma=0.1)

完整代码

Reference

tranfer learning cs231n notes
Transfer learning for computer vision turorial
Pytorchtutorial-transfer-learning