PyTorch基础（六）迁移学习

在实际工程中，基本没有人会从零开始（随机初始化）训练一个完整的卷积网络，因为相对于网络，很难得到一个足够大的数据集（网络很深，需要足够大数据集）。通常的做法是在一个很大的数据集上进行预训练得到卷积网络ConvNet，然后将这个ConvNet的参数作为目标任务的初始化参数或者固定这些参数。

迁移学习的两个使用场景

• 微调（Fine-tune）Convnet：使用预训练的网络（如在imagenet 1000上训练而来的网络）来初始化自己的网络，而不是随机初始化。其他的训练步骤不变。
• 将Convnet看成固定的特征提取器 ：首先固定ConvNet处理最后的全连接层外的其他所有层。最后的全连接层被替换成一个新的随机初始化的层，只有这个新的层会被训练【只有这层的参数会在反向传播时更新】

下面是利用PyTorch进行迁移学习的步骤，要解决的问题是训练一个模型来对蚂蚁和蜜蜂进行分类。

#!/usr/bin/env torch
# -*- coding:utf-8 -*-
# @Time  : 2021/2/7, 14:08
# @Author: Lee
# @File  : classify_ants_bees.py

# 导入相关的包
from __future__ import print_function, division

import torch
import torch.nn as nn
import torch.optim as optim
from torch.optim import lr_scheduler
import numpy as np
import torchvision
from torchvision import datasets, models, transforms
import matplotlib.pyplot as plt
import time
import os
import copy

plt.ion()  # plt绘图使用交互模式，而不是默认的阻塞模式，绘制第一个figure后继续执行后面的程序，可绘制多张figure

# 加载数据
# 训练集数据扩充和归一化， 在验证集上仅需归一化
data_transforms = {
    'train': transforms.Compose([
        transforms.RandomResizedCrop(224),  # 随机剪裁一个area然后再resize
        transforms.RandomHorizontalFlip(),  # 随机水平翻转
        transforms.ToTensor(),
        transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.2225])
    ]),
    'val': transforms.Compose([
        transforms.Resize(256),
        transforms.CenterCrop(224),
        transforms.ToTensor(),
        transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.2225])
    ]),
}

data_dir = 'data/hymenoptera_data'
image_datasets = {x: datasets.ImageFolder(os.path.join(data_dir, x),
                                           data_transforms[x])
                  for x in ['train', 'val']}
dataloaders = {x: torch.utils.data.DataLoader(image_datasets[x], batch_size=4,
                                               shuffle=True, num_workers=4)
               for x in ['train', 'val']}
dataset_sizes = {x: len(image_datasets[x]) for x in ['train', 'val']}
class_name = image_datasets['train'].classes

device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")


# 可视化部分图像数据，以便了解数据扩充
def imshow(inp, title=None):
    """
    Imshow for tensor
    """
    inp = inp.numpy().transpose((1, 2, 0))
    mean = np.array([0.485, 0.456, 0.406])
    std = np.array([0.229, 0.224, 0.2225])
    inp = std * inp + mean
    inp = np.clip(inp, 0, 1)
    plt.imshow(inp)
    if title is not None:
        plt.title(title)
    plt.pause(0.001)  # pause a bit so that plots are updated


# 训练模型，编写一个通用函数来训练模型
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):
        print('Epoch {}/{}'.format(epoch, num_epochs - 1))
        print('-' * 10)

        # 每个epoch都有一个训练和验证阶段
        for phase in ['train', 'val']:
            if phase == 'train':
                # scheduler.step()
                model.train()  # Set model to train mode
            else:
                model.eval()  # Set model to evaluate mode

            running_loss = 0.0
            running_corrects = 0

            # 迭代数据
            for inputs, labels in dataloaders[phase]:
                inputs = inputs.to(device)
                labels = labels.to(device)

                # 零参数梯度
                optimizer.zero_grad()

                # 前向传播 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)

                    # 后向传播，尽在训练集阶段优化
                    if phase == 'train':
                        loss.backward()
                        optimizer.step()

                # 统计
                running_loss += loss.item() * inputs.size(0)
                running_corrects += torch.sum(preds == labels.data)

            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))

            # 深度拷贝
            if phase == 'val' and epoch_acc > best_acc:
                best_acc = epoch_acc
                best_model_wts = copy.deepcopy(model.state_dict())

        if phase == 'train':
            scheduler.step()

        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))

    # 加载最佳模型权重
    model.load_state_dict(best_model_wts)
    return model


# 可视化模型的预测结果
def visualize_model(model, num_images=6):
    was_training = model.training
    model.eval()
    image_so_far = 0
    fig = plt.figure()

    with torch.no_grad():
        for i ,(inputs, labels) in enumerate(dataloaders['val']):
            inputs = inputs.yo(device)
            labels = labels.to(device)

            outputs = model(inputs)
            _, preds = torch.max(outputs, 1)

            for j in range(inputs.size()[0]):
                image_so_far += 1
                ax = plt.subplot(num_images // 2, 3, image_so_far)
                ax.axis('off')
                ax.set_title('predicted: {}'.format(class_name[preds[j]]))
                imshow(inputs.cpu().data[j])

                if image_so_far == num_images:
                    model.train(mode=was_training)
                    return
        model.train(mode=was_training)


if __name__ == '__main__':
    pass

定义好相关的函数后，在以下各种情况下调用函数，首先，先可视化部分训练图像，以便充分了解数据扩充，其主函数如下：

if __name__ == '__main__':
    # 获取一批训练数据
    inputs, classes = next(iter(dataloaders['train']))
    # 批量制作网络
    out = torchvision.utils.make_grid(inputs)
    imshow(out, title=[class_name[x] for x in classes])
    plt.pause(0)  # 避免图像一闪而过

运行结果如下：

然后，微调ConvNet，加载预训练模型并重置最终全连接层，并进行训练：其主函数如下：

if __name__=='__main__': 
    model_ft = models.resnet18(pretrained=True)
    num_ftrs = model_ft.fc.in_features
    model_ft.fc = nn.Linear(num_ftrs, 2)

    model_ft = model_ft.to(device)
    criterion = nn.CrossEntropyLoss()

    # 观察所有参数都正在优化
    optimizier_ft = optim.SGD(model_ft.parameters(), lr=0.001, momentum=0.9)
    # 每7个epochs衰减LR通过设置gamma=0.1
    exp_lr_scheduler = lr_scheduler.StepLR(optimizier_ft, step_size=7, gamma=0.1)
    model_ft = train_model(model_ft, criterion, optimizier_ft, exp_lr_scheduler,     
               num_epochs=25)
    # 保存模型
    torch.save(model_ft, 'models/model')  # 此处暂时以保存整个模型结构和参数为例

训练过程中打印如下（省略部分）：

Epoch 0/24
----------
train Loss: 0.5123 Acc: 0.7172
val Loss: 0.2224 Acc: 0.9281

Epoch 1/24
----------
train Loss: 0.4808 Acc: 0.8033
val Loss: 0.2503 Acc: 0.9020

...


Epoch 19/24
----------
train Loss: 0.3454 Acc: 0.8770
val Loss: 0.1923 Acc: 0.9542

...

Epoch 23/24
----------
train Loss: 0.2384 Acc: 0.8975
val Loss: 0.2453 Acc: 0.9150

Epoch 24/24
----------
train Loss: 0.3968 Acc: 0.8566
val Loss: 0.3225 Acc: 0.9085

Training complete in 37m 15s
Best val Acc: 0.9542

再调用上面训练好的模型，并做可视化预测，代码如下：

    model_ft = torch.load('models/model')  # 加载模型
    visualize_model(model_ft)
    plt.pause(0)

运行得到如下结果：

然后，再将ConvNet作为固定特征提取器，在这里需要冻结除最后一层之外的所有网络。通过设置requires_grad=Falsebackward()来冻结参数，这样反向传播backward()的时候它们的梯度就不会被计算。主函数代码如下：

if __name__ == '__main__':
    # ConvNet作为特征提取器
    """
    在这里需要冻结最后一层之外的所有网络。通过设置requires_grad = Falsebackward()来冻结
    参数，这样在反向传播backward()的时候它们的梯度就不会被计算
    """
    model_conv = torchvision.models.resnet18(pretrained=True)
    for param in model_conv.parameters():
        param.requires_grad = False

    # Parameters of newly constructed modules have requires_grad = True by default
    num_ftrs = model_conv.fc.in_features
    model_conv.fc = nn.Linear(num_ftrs, 2)
    model_conv = model_conv.to(device)
    criterion = nn.CrossEntropyLoss()

    # Oberseve that parameters of final layer are being optimized as opposed to before
    optimizer_conv = optim.SGD(model_conv.fc.parameters(), lr=0.001, momentum=0.9)
    # Decay LR by a factor of 0.1 every 7 epochs
    exp_lr_scheduler = lr_scheduler.StepLR(optimizer_conv, step_size=7, gamma=0.1)
    model_conv = train_model(model_conv, criterion, optimizer_conv, exp_lr_scheduler,                     
                  num_epochs=25)
    torch.save(model_conv, 'models/model_conv')

运行该程序，训练过程打印如下（有部分省略）：

Epoch 0/24
----------
train Loss: 0.6712 Acc: 0.6434
val Loss: 0.2472 Acc: 0.9150

Epoch 1/24
----------
train Loss: 0.5104 Acc: 0.7664
val Loss: 0.1941 Acc: 0.9346

...

Epoch 24/24
----------
train Loss: 0.2855 Acc: 0.8484
val Loss: 0.2148 Acc: 0.9477

Training complete in 21m 37s
Best val Acc: 0.9477

再基于上面重新训练的模型，进行模型效果可视化，将主函数换为如下代码:

# 模型评估效果可视化
    model_conv = torch.load('models/model_conv')  # 加载模型
    visualize_model(model_conv)
    plt.pause(0)

运行结果如下：

附：本文中学习率衰减用的SetpLR，关于学习率衰减可查看https://zhuanlan.zhihu.com/p/93624972。