TRANSFER LEARNING TUTORIAL

在本教程中，您将学习如何使用迁移学习来训练您的网络。 您可以在cs231n笔记上阅读有关迁移学习的更多信息。

引用这些笔记，

在实践中，很少有人从头开始训练整个卷积网络（随机初始化），因为拥有足够大小的数据集是相对罕见的。相反，通常在非常大的数据集（例如ImageNet，其包含具有1000个类别的120万个图像）上预先训练ConvNet，然后使用ConvNet作为感兴趣的任务的初始化或固定特征提取器。

这两个主要的迁移学习场景如下：

• Finetuning the convnet：我们使用预训练网络初始化网络，而不是随机初始化，就像在imagenet 1000数据集上训练的网络一样。 其余训练看起来像往常一样。
• ConvNet作为固定特征提取器：在这里，我们将冻结除最终完全连接层之外的所有网络的权重。 最后一个完全连接的层被替换为具有随机权重的新层，并且仅训练该层。

# 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

Load Data

我们将使用torchvision和torch.utils.data包来加载数据。

我们今天要解决的问题是训练一个模型来对蚂蚁和蜜蜂进行分类。 我们有大约120个训练图像，每个图像用于蚂蚁和蜜蜂。 每个类有75个验证图像。 通常，如果从头开始训练，这是一个非常小的数据集。 由于我们正在使用迁移学习，我们应该能够合理地推广。

该数据集是imagenet的一个非常小的子集。

从此处下载数据并将其解压缩到当前目录。

# Data augmentation and normalization for training
# Just normalization for validation
data_transforms = {
    'train': transforms.Compose([
        transforms.RandomResizedCrop(224),
        transforms.RandomHorizontalFlip(),
        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")

Visualize a few images

让我们可视化一些训练图像，以便了解数据增强。

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


# Get a batch of training data
inputs, classes = next(iter(dataloaders['train']))

# Make a grid from batch
out = torchvision.utils.make_grid(inputs)

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

Training the model

现在，让我们编写一个通用函数来训练模型。 在这里，我们将说明：

• 安排学习率
• 保存最好的模型

在下文中，参数scheduler是来自torch.optim.lr_scheduler的LR调度程序对象。

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)

        # Each epoch has a training and validation phase
        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

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

                # zero the parameter gradients
                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)

                    # backward + optimize only if in training phase
                    if phase == 'train':
                        loss.backward()
                        optimizer.step()

                # statistics
                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))

            # 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

Visualizing the model predictions

用于显示少量图像预测的通用函数

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)

Finetuning the convnet

加载预训练模型并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()

# Observe that all parameters are being optimized
optimizer_ft = optim.SGD(model_ft.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_ft, step_size=7, gamma=0.1)

Train and evaluate

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)

Out:

Epoch 0/24
----------
train Loss: 0.6223 Acc: 0.6516
val Loss: 0.3521 Acc: 0.8889

Epoch 1/24
----------
train Loss: 0.6196 Acc: 0.7582
val Loss: 0.2696 Acc: 0.9020

Epoch 2/24
----------
train Loss: 0.4888 Acc: 0.7992
val Loss: 0.3266 Acc: 0.8693

Epoch 3/24
----------
train Loss: 0.4684 Acc: 0.8115
val Loss: 0.2606 Acc: 0.9281

Epoch 4/24
----------
train Loss: 0.7273 Acc: 0.7582
val Loss: 0.6259 Acc: 0.8170

Epoch 5/24
----------
train Loss: 0.4543 Acc: 0.8279
val Loss: 0.3323 Acc: 0.8954

Epoch 6/24
----------
train Loss: 0.4281 Acc: 0.8238
val Loss: 0.4571 Acc: 0.8824

Epoch 7/24
----------
train Loss: 0.5758 Acc: 0.8443
val Loss: 0.3238 Acc: 0.9150

Epoch 8/24
----------
train Loss: 0.3818 Acc: 0.8361
val Loss: 0.2771 Acc: 0.9085

Epoch 9/24
----------
train Loss: 0.3732 Acc: 0.8402
val Loss: 0.3123 Acc: 0.8889

Epoch 10/24
----------
train Loss: 0.2966 Acc: 0.8566
val Loss: 0.2823 Acc: 0.9150

Epoch 11/24
----------
train Loss: 0.3123 Acc: 0.8566
val Loss: 0.2809 Acc: 0.9085

Epoch 12/24
----------
train Loss: 0.2116 Acc: 0.9098
val Loss: 0.2842 Acc: 0.9085

Epoch 13/24
----------
train Loss: 0.3401 Acc: 0.8525
val Loss: 0.2774 Acc: 0.9085

Epoch 14/24
----------
train Loss: 0.2864 Acc: 0.8811
val Loss: 0.2774 Acc: 0.9085

Epoch 15/24
----------
train Loss: 0.2472 Acc: 0.8811
val Loss: 0.2796 Acc: 0.8954

Epoch 16/24
----------
train Loss: 0.3174 Acc: 0.8770
val Loss: 0.2770 Acc: 0.9150

Epoch 17/24
----------
train Loss: 0.2507 Acc: 0.8975
val Loss: 0.2726 Acc: 0.8954

Epoch 18/24
----------
train Loss: 0.3065 Acc: 0.8689
val Loss: 0.2660 Acc: 0.8954

Epoch 19/24
----------
train Loss: 0.3193 Acc: 0.8730
val Loss: 0.2824 Acc: 0.9020

Epoch 20/24
----------
train Loss: 0.3228 Acc: 0.8525
val Loss: 0.2788 Acc: 0.9085

Epoch 21/24
----------
train Loss: 0.2769 Acc: 0.9057
val Loss: 0.2679 Acc: 0.9150

Epoch 22/24
----------
train Loss: 0.2798 Acc: 0.8730
val Loss: 0.2728 Acc: 0.9085

Epoch 23/24
----------
train Loss: 0.2723 Acc: 0.8893
val Loss: 0.2569 Acc: 0.9085

Epoch 24/24
----------
train Loss: 0.2465 Acc: 0.8893
val Loss: 0.2652 Acc: 0.9150

Training complete in 1m 8s
Best val Acc: 0.928105

visualize_model(model_ft)

ConvNet as fixed feature extractor

在这里，我们需要冻结除最后一层之外的所有网络。 我们需要设置requires_grad == False来冻结参数，以便不在backward()中计算梯度。

您可以在此处的文档中阅读更多相关信息。

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)

Train and evaluate

在CPU上，与前一个场景相比，这将花费大约一半的时间。 这是预期的，因为不需要为大多数网络计算梯度。 但是，需要计算梯度。

model_conv = train_model(model_conv, criterion, optimizer_conv,
                         exp_lr_scheduler, num_epochs=25)

Out:

Epoch 0/24
----------
train Loss: 0.6739 Acc: 0.6393
val Loss: 0.3576 Acc: 0.8301

Epoch 1/24
----------
train Loss: 0.5562 Acc: 0.7295
val Loss: 0.2005 Acc: 0.9281

Epoch 2/24
----------
train Loss: 0.7259 Acc: 0.6967
val Loss: 0.5577 Acc: 0.7712

Epoch 3/24
----------
train Loss: 0.6574 Acc: 0.7582
val Loss: 0.1995 Acc: 0.9281

Epoch 4/24
----------
train Loss: 0.4485 Acc: 0.8443
val Loss: 0.2090 Acc: 0.9412

Epoch 5/24
----------
train Loss: 0.5107 Acc: 0.7951
val Loss: 0.2062 Acc: 0.9346

Epoch 6/24
----------
train Loss: 0.5821 Acc: 0.8156
val Loss: 0.1935 Acc: 0.9542

Epoch 7/24
----------
train Loss: 0.3424 Acc: 0.8484
val Loss: 0.2099 Acc: 0.9477

Epoch 8/24
----------
train Loss: 0.3542 Acc: 0.8484
val Loss: 0.2294 Acc: 0.9346

Epoch 9/24
----------
train Loss: 0.3682 Acc: 0.8361
val Loss: 0.1848 Acc: 0.9477

Epoch 10/24
----------
train Loss: 0.3868 Acc: 0.8361
val Loss: 0.2021 Acc: 0.9477

Epoch 11/24
----------
train Loss: 0.3017 Acc: 0.8730
val Loss: 0.1896 Acc: 0.9608

Epoch 12/24
----------
train Loss: 0.4049 Acc: 0.8361
val Loss: 0.1831 Acc: 0.9477

Epoch 13/24
----------
train Loss: 0.3673 Acc: 0.8361
val Loss: 0.1891 Acc: 0.9542

Epoch 14/24
----------
train Loss: 0.3887 Acc: 0.8320
val Loss: 0.1927 Acc: 0.9477

Epoch 15/24
----------
train Loss: 0.4134 Acc: 0.8156
val Loss: 0.1914 Acc: 0.9542

Epoch 16/24
----------
train Loss: 0.3637 Acc: 0.8689
val Loss: 0.2104 Acc: 0.9412

Epoch 17/24
----------
train Loss: 0.3139 Acc: 0.8484
val Loss: 0.1999 Acc: 0.9608

Epoch 18/24
----------
train Loss: 0.2633 Acc: 0.8893
val Loss: 0.2157 Acc: 0.9412

Epoch 19/24
----------
train Loss: 0.4095 Acc: 0.8320
val Loss: 0.2008 Acc: 0.9412

Epoch 20/24
----------
train Loss: 0.2604 Acc: 0.9016
val Loss: 0.2012 Acc: 0.9542

Epoch 21/24
----------
train Loss: 0.3425 Acc: 0.8607
val Loss: 0.2186 Acc: 0.9477

Epoch 22/24
----------
train Loss: 0.3245 Acc: 0.8607
val Loss: 0.1831 Acc: 0.9477

Epoch 23/24
----------
train Loss: 0.3047 Acc: 0.8607
val Loss: 0.2010 Acc: 0.9542

Epoch 24/24
----------
train Loss: 0.2519 Acc: 0.8975
val Loss: 0.2045 Acc: 0.9477

Training complete in 0m 35s
Best val Acc: 0.960784

visualize_model(model_conv)

plt.ioff()
plt.show()