PyTorch迁移学习

实际中，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。

转移学习的两个主要场景：

• 微调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。其它的训练步骤不变。
• 将Convnet看成固定的特征提取器: 首先固定ConvNet，除了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[只有这层参数会在反向传播时更新]

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

1.导入相关的包

# License: BSD

# Author: Sasank Chilamkurthy

 

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()   # interactive mode

2.加载数据

要解决的问题是，训练一个模型来分类蚂蚁ants和蜜蜂bees。ants和bees各有约120张训练图片。每个类有75张验证图片。从零开始，在如此小的数据集上进行训练，通常是很难泛化的。由于使用迁移学习，模型的泛化能力会相当好。该数据集是imagenet的一个非常小的子集。下载数据，并将其解压缩到当前目录。

＃训练集数据扩充和归一化

＃在验证集上仅需要归一化

data_transforms = {

    'train': transforms.Compose([

        transforms.RandomResizedCrop(224), #随机裁剪一个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, 0.225])

    ]),

    '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])

    ]),

}

 

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_names = image_datasets['train'].classes

 

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

3.可视化部分图像数据

可视化部分训练图像，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, 0.225])

    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

 

 

# 获取一批训练数据

inputs, classes = next(iter(dataloaders['train']))

 

# 批量制作网格

out = torchvision.utils.make_grid(inputs)

 

imshow(out, title=[class_names[x] for x in classes])

 

4.训练模型

编写一个通用函数来训练模型。下面将说明： * 调整学习速率 * 保存最好的模型

下面的参数scheduler，是一个来自 torch.optim.lr_scheduler的学习速率调整类的对象(LR scheduler object)。

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 training 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

                # 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+仅在训练阶段进行优化

                    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))

 

            # 深度复制mo

            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))

 

    # 加载最佳模型权重

    model.load_state_dict(best_model_wts)

    return model

5.可视化模型的预测结果

#一个通用的展示少量预测图片的函数

def visualize_model(model, num_images=6):

    was_training = model.training

    model.eval()

    images_so_far = 0

    fig = plt.figure()

 

    with torch.no_grad():

        for i, (inputs, labels) in enumerate(dataloaders['val']):

            inputs = inputs.to(device)

            labels = labels.to(device)

 

            outputs = model(inputs)

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

 

            for j in range(inputs.size()[0]):

                images_so_far += 1

                ax = plt.subplot(num_images//2, 2, images_so_far)

                ax.axis('off')

                ax.set_title('predicted: {}'.format(class_names[preds[j]]))

                imshow(inputs.cpu().data[j])

 

                if images_so_far == num_images:

                    model.train(mode=was_training)

                    return

        model.train(mode=was_training)

6.场景1：微调ConvNet

加载预训练模型，重置最终完全连接的图层。

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

optimizer_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(optimizer_ft, step_size=7, gamma=0.1)

训练和评估模型

（1）训练模型 该过程在CPU上，需要大约15-25分钟，但是在GPU上，它只需不到一分钟。

model_ft = train_model(model_ft, criterion, optimizer_ft, exp_lr_scheduler,

                       num_epochs=25)

• 输出

Epoch 0/24

----------

train Loss: 0.7032 Acc: 0.6025

val Loss: 0.1698 Acc: 0.9412

 

Epoch 1/24

----------

train Loss: 0.6411 Acc: 0.7787

val Loss: 0.1981 Acc: 0.9281

·

·

·

Epoch 24/24

----------

train Loss: 0.2812 Acc: 0.8730

val Loss: 0.2647 Acc: 0.9150

 

Training complete in 1m 7s

Best val Acc: 0.941176

（2）模型评估效果可视化

visualize_model(model_ft)

• 输出
•  

   

   

7.场景2：ConvNet作为固定特征提取器

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，这样在反向传播backward()的时候，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()

 

# Observe that only 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)

训练和评估

（1）训练模型 在CPU上，与前一个场景相比，这将花费大约一半的时间，因为不需要为大多数网络计算梯度。但需要计算转发。

model_conv = train_model(model_conv, criterion, optimizer_conv,

                         exp_lr_scheduler, num_epochs=25)

• 输出

Epoch 0/24

----------

train Loss: 0.6400 Acc: 0.6434

val Loss: 0.2539 Acc: 0.9085

·

·

·

Epoch 23/24

----------

train Loss: 0.2988 Acc: 0.8607

val Loss: 0.2151 Acc: 0.9412

 

Epoch 24/24

----------

train Loss: 0.3519 Acc: 0.8484

val Loss: 0.2045 Acc: 0.9412

 

Training complete in 0m 35s

Best val Acc: 0.954248

（2）模型评估效果可视化

visualize_model(model_conv)

 

plt.ioff()

plt.show()

• 输出 

 

8.文件下载

• py文件
• jupyter文件