这篇文章翻译自 Pytorch 官方教程 Transfer Learning Tutorial
原作者:Sasank Chilamkurthy
Note: 点击这里下载完整示例代码
在这篇教程中,你将会学到如何利用迁移学习来训练你的网络。你可以通过 cs231n notes 了解更多关于迁移学习的信息。
引用 cs231n notes 中的一段话
在实践中,很少有人会从头开始训练一个卷积神经网络(随机初始化),因为你很难拥有一个足够大的数据集。事实上,更常见的做法是先在一个非常大的数据集(比如 ImageNet,该数据集含有涵盖了 1000 个类别的 120 万张图片)上预训练一个卷积神经网络,然后利用该网络中参数作为初始参数,或者把该网络当作另一项任务的固定特征提取器。
迁移学习主要在以下两个场景下使用:
# 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() # 交互模式
我们将使用 torchvision
和 torch.utils.data
两个 packages 来读取数据。
我们今天的目标是建立一个可以分辨蚂蚁和蜜蜂的分类器,但是我们只有蚂蚁和蜜蜂的图片各约 120 张用于训练,75 张用于验证集。通常来说,如果要从头训练一个模型,这个数据集是非常小的。因此我们要利用迁移学习。
这个数据集是 ImageNet 的一个很小的子集。
Note: 从 这里 下载数据并将其解压到当前文件夹。
# 对训练集使用 data augmentation 和 normalization
# 对验证集只使用 normalization
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")
为了理解 data augmentation,我们来看看一些图像。
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
# 获取训练数据中的一个 batch
inputs, classes = next(iter(dataloaders['train']))
out = torchvision.utils.make_grid(inputs)
imshow(out, title=[class_names[x] for x in classes])
现在,为了训练模型,我们应该写一些通用函数。在这里我们将阐述以下两点
下面参数中的 scheduler
是torch.optim.lr_scheduler
包中的 LR scheduler 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)
# 每次遍历都要经过训练集和验证集
for phase in ['train', 'val']:
if phase == 'train':
scheduler.step()
model.train() # 设置模型为训练模式
else:
model.eval() # 设置模型为验证模式
running_loss = 0.0
running_corrects = 0
# 迭代
for inputs, labels in dataloaders[phase]:
inputs = inputs.to(device)
labels = labels.to(device)
# 清零梯度
optimizer.zero_grad()
# 前向传播
# 只在训练时计算梯度
with torch.set_grad_enabled(phase == 'train'):
outputs = model(inputs)
_, preds = torch.max(outputs, 1)
loss = criterion(outputs, labels)
# 只有在训练时才进行反向传播和参数更新
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))
# 找到最好的模型
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))
# 读取最好模型
model.load_state_dict(best_model_wts)
return model
这是一个用于展示预测结果的通用函数
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)
读取一个预训练网络并重置最后的全连接层
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()
# 这里所有参数都会更新
optimizer_ft = optim.SGD(model_ft.parameters(), lr=0.001, momentum=0.9)
# 学习率每 7 次迭代以 0.1 为因子衰减
exp_lr_scheduler = lr_scheduler.StepLR(optimizer_ft, step_size=7, gamma=0.1)
在 CPU 上训练会花费大约 15-25 分钟,而在 GPU 上则要不了一分钟。
model_ft = train_model(model_ft, criterion, optimizer_ft, exp_lr_scheduler,
num_> Epochs=25)
输出:
Epoch 0/24
——————
train Loss: 0.5900 Acc: 0.7131
val Loss: 0.2508 Acc: 0.9020
Epoch 1/24
——————
train Loss: 0.6034 Acc: 0.7828
val Loss: 0.3181 Acc: 0.8627
Epoch 2/24
——————
train Loss: 0.6150 Acc: 0.7582
val Loss: 0.4903 Acc: 0.8366
Epoch 3/24
——————
train Loss: 0.6650 Acc: 0.7377
val Loss: 0.6294 Acc: 0.7582
Epoch 4/24
——————
train Loss: 0.4935 Acc: 0.7828
val Loss: 0.2644 Acc: 0.8889
Epoch 5/24
——————
train Loss: 0.3841 Acc: 0.8238
val Loss: 0.24 08 Acc: 0.9216
Epoch 6/24
——————
train Loss: 0.5352 Acc: 0.8156
val Loss: 0.2250 Acc: 0.9150
Epoch 7/24
——————
train Loss: 0.2252 Acc: 0.9385
val Loss: 0.1917 Acc: 0.9477
Epoch 8/24
——————
train Loss: 0.3395 Acc: 0.8197
val Loss: 0.1738 Acc: 0.9477
Epoch 9/24
——————
train Loss: 0.3363 Acc: 0.8607
val Loss: 0.2522 Acc: 0.9216
Epoch 10/24
——————
train Loss: 0.2878 Acc: 0.8607
val Loss: 0.1787 Acc: 0.9412
Epoch 11/24
——————
train Loss: 0.2831 Acc: 0.8770
val Loss: 0.1805 Acc: 0.9346
Epoch 12/24
——————
train Loss: 0.2290 Acc: 0.9016
val Loss: 0.1898 Acc: 0.9412
Epoch 13/24
——————
train Loss: 0.24 94 Acc: 0.9016
val Loss: 0.1729 Acc: 0.9412
Epoch 14/24
——————
train Loss: 0.3435 Acc: 0.8689
val Loss: 0.1736 Acc: 0.9412
Epoch 15/24
——————
train Loss: 0.2274 Acc: 0.9057
val Loss: 0.1692 Acc: 0.9542
Epoch 16/24
——————
train Loss: 0.3154 Acc: 0.8689
val Loss: 0.1742 Acc: 0.9412
Epoch 17/24
——————
train Loss: 0.2749 Acc: 0.8893
val Loss: 0.1826 Acc: 0.9412
Epoch 18/24
——————
train Loss: 0.2673 Acc: 0.8770
val Loss: 0.1731 Acc: 0.9281
Epoch 19/24
——————
train Loss: 0.2865 Acc: 0.8730
val Loss: 0.1867 Acc: 0.9346
Epoch 20/24
——————
train Loss: 0.3061 Acc: 0.8648
val Loss: 0.1966 Acc: 0.9477
Epoch 21/24
——————
train Loss: 0.2638 Acc: 0.9016
val Loss: 0.1973 Acc: 0.9477
Epoch 22/24
——————
train Loss: 0.2602 Acc: 0.8893
val Loss: 0.1769 Acc: 0.9281
Epoch 23/24
——————
train Loss: 0.2817 Acc: 0.9016
val Loss: 0.1756 Acc: 0.9412
Epoch 24 /24
——————
train Loss: 0.2959 Acc: 0.8730
val Loss: 0.1790 Acc: 0.9281
Training complete in 1m 8s
Best val Acc: 0.95424 8
visualize_model(model_ft)
现在,除了最后的全连接层,我们要冻结网络中其余部分的所有参数。我们使用 requires_grad = False
来冻结参数,bachward()
便不会计算这些参数的梯度。
你可以在 这里 读到更多信息。
model_conv = torchvision.models.resnet18(pretrained=True)
for param in model_conv.parameters():
param.requires_grad = False
# 新构建模块中的参数的 requires_grad 默认为 True
num_ftrs = model_conv.fc.in_features
model_conv.fc = nn.Linear(num_ftrs, 2)
model_conv = model_conv.to(device)
criterion = nn.CrossEntropyLoss()
# 现在只有最后的全连接层的参数会更新
optimizer_conv = optim.SGD(model_conv.fc.parameters(), lr=0.001, momentum=0.9)
# 学习率每 7 次迭代以 0.1 为因子衰减
exp_lr_scheduler = lr_scheduler.StepLR(optimizer_conv, step_size=7, gamma=0.1)
在 CPU 上,这会花费约之前一半的时间,因为大多数参数的梯度不用计算了,不过这些参数仍然会参与前向传播。
model_conv = train_model(model_conv, criterion, optimizer_conv,
exp_lr_scheduler, num_> Epochs=25)
输出:
Epoch 0/24
——————
train Loss: 0.6463 Acc: 0.6803
val Loss: 0.1949 Acc: 0.9477
Epoch 1/24
——————
train Loss: 0.4923 Acc: 0.8033
val Loss: 0.1696 Acc: 0.9477
Epoch 2/24
——————
train Loss: 0.4234 Acc: 0.8115
val Loss: 0.4379 Acc: 0.7712
Epoch 3/24
——————
train Loss: 0.5606 Acc: 0.7582
val Loss: 0.6383 Acc: 0.7451
Epoch 4/24
——————
train Loss: 0.7560 Acc: 0.7295
val Loss: 0.1888 Acc: 0.9412
Epoch 5/24
——————
train Loss: 0.4316 Acc: 0.8197
val Loss: 0.1999 Acc: 0.9477
Epoch 6/24
——————
train Loss: 0.7722 Acc: 0.7131
val Loss: 0.1975 Acc: 0.9477
Epoch 7/24
——————
train Loss: 0.3685 Acc: 0.8607
val Loss: 0.2000 Acc: 0.9477
Epoch 8/24
——————
train Loss: 0.2968 Acc: 0.8811
val Loss: 0.1916 Acc: 0.9477
Epoch 9/24
——————
train Loss: 0.3396 Acc: 0.8525
val Loss: 0.2165 Acc: 0.9542
Epoch 10/24
——————
train Loss: 0.3885 Acc: 0.8320
val Loss: 0.2109 Acc: 0.9542
Epoch 11/24
——————
train Loss: 0.4107 Acc: 0.8156
val Loss: 0.1881 Acc: 0.9477
Epoch 12/24
——————
train Loss: 0.3249 Acc: 0.8730
val Loss: 0.1747 Acc: 0.9542
Epoch 13/24
——————
train Loss: 0.3439 Acc: 0.8525
val Loss: 0.1950 Acc: 0.9477
Epoch 14/24
——————
train Loss: 0.3641 Acc: 0.8443
val Loss: 0.1992 Acc: 0.9412
Epoch 15/24
——————
train Loss: 0.3272 Acc: 0.8443
val Loss: 0.2320 Acc: 0.9412
Epoch 16/24
——————
train Loss: 0.3102 Acc: 0.8730
val Loss: 0.1867 Acc: 0.9477
Epoch 17/24
——————
train Loss: 0.4226 Acc: 0.8238
val Loss: 0.1872 Acc: 0.9542
Epoch 18/24
——————
train Loss: 0.3452 Acc: 0.8443
val Loss: 0.1812 Acc: 0.9542
Epoch 19/24
——————
train Loss: 0.3697 Acc: 0.8525
val Loss: 0.1890 Acc: 0.9477
Epoch 20/24
——————
train Loss: 0.3078 Acc: 0.8607
val Loss: 0.1976 Acc: 0.9608
Epoch 21/24
——————
train Loss: 0.3161 Acc: 0.8770
val Loss: 0.1982 Acc: 0.9412
Epoch 22/24
——————
train Loss: 0.3749 Acc: 0.8320
val Loss: 0.2035 Acc: 0.9477
Epoch 23/24
——————
train Loss: 0.3298 Acc: 0.8525
val Loss: 0.1855 Acc: 0.9477
Epoch 24/24
——————
train Loss: 0.3597 Acc: 0.8402
val Loss: 0.1878 Acc: 0.9542
Training complete in 0m 34s
Best val Acc: 0.960784
visualize_model(model_conv)
plt.ioff()
plt.show()
下载 Python 源代码:transfer_learning_tutorial.py
下载 Jupyter Notebook: transfer_learning_tutorial.ipynb