前言
在訓(xùn)練深度學(xué)習(xí)模型時(shí)，有時(shí)候我們沒(méi)有海量的訓(xùn)練樣本，只有少數(shù)的訓(xùn)練樣本(比如幾百個(gè)圖片)诲宇，幾百個(gè)訓(xùn)練樣本顯然對(duì)于深度學(xué)習(xí)遠(yuǎn)遠(yuǎn)不夠。這時(shí)候惶翻，我們可以使用別人預(yù)訓(xùn)練好的網(wǎng)絡(luò)模型權(quán)重姑蓝，在此基礎(chǔ)上進(jìn)行訓(xùn)練，這就引入了一個(gè)概念——遷移學(xué)習(xí)(Transfer Learning)吕粗。
遷移學(xué)習(xí)
What(什么是遷移學(xué)習(xí))
遷移學(xué)習(xí)（Transfer Learning,TL）對(duì)于人類來(lái)說(shuō)纺荧，就是掌握舉一反三的學(xué)習(xí)能力。比如我們學(xué)會(huì)騎自行車后颅筋，學(xué)騎摩托車就很簡(jiǎn)單了宙暇；在學(xué)會(huì)打羽毛球之后，再學(xué)打網(wǎng)球也就沒(méi)那么難了议泵。對(duì)于計(jì)算機(jī)而言客给，所謂遷移學(xué)習(xí)，就是能讓現(xiàn)有的模型算法稍加調(diào)整即可應(yīng)用于一個(gè)新的領(lǐng)域和功能的一項(xiàng)技術(shù)
How(如何進(jìn)行遷移學(xué)習(xí))
首先需要選擇一個(gè)預(yù)訓(xùn)練好的模型肢簿，需要注意的是該模型的訓(xùn)練過(guò)程最好與我們要進(jìn)行訓(xùn)練的任務(wù)相似靶剑。比如我們要訓(xùn)練一個(gè)Cat,dog圖像分類的模型，最好應(yīng)該選擇一個(gè)圖像分類的預(yù)訓(xùn)練模型池充。
針對(duì)實(shí)際任務(wù)桩引，對(duì)網(wǎng)絡(luò)結(jié)構(gòu)進(jìn)行調(diào)整。比如找到了一個(gè)預(yù)訓(xùn)練好的AlexNet（1000類別）收夸， 但是我們實(shí)際的任務(wù)的2分類坑匠，因此需要把最后一層的全連接輸出改為2.
Why(為何要使用遷移學(xué)習(xí))
https://pytorch.org/tutorials/beginner/transfer_learning_tutorial.html
In practice, very few people train an entire Convolutional Network from scratch (with random initialization), because it is relatively rare to have a dataset of sufficient size. Instead, it is common to pretrain a ConvNet on a very large dataset (e.g. ImageNet, which contains 1.2 million images with 1000 categories), and then use the ConvNet either as an initialization or a fixed feature extractor for the task of interest.
目的
- 了解ResNet
- 基于預(yù)訓(xùn)練好的ResNet-18, 進(jìn)行一個(gè)圖像二分類遷移學(xué)習(xí)
開發(fā)/測(cè)試環(huán)境
- Ubuntu 18.04
- pycharm
- Anaconda3, python3.6
- pytorch1.0, torchvision
ResNet-18
image.png
實(shí)驗(yàn)內(nèi)容
準(zhǔn)備數(shù)據(jù)集
- 訓(xùn)練集合
- 驗(yàn)證集合
下載好之后，復(fù)制到工程 /data/ 路徑下
訓(xùn)練集合卧惜，驗(yàn)證集合
訓(xùn)練集厘灼，驗(yàn)證集 分別包含2個(gè)子文件夾，這是一個(gè)2分類問(wèn)題咽瓷。分類對(duì)象：螞蟻设凹，蜜蜂
-
代碼
因?yàn)橛?xùn)練一個(gè)2分類的模型，數(shù)據(jù)集加載直接使用pytorch提供的API——ImageFolder最方便茅姜。原始圖像為jpg格式闪朱，在制作數(shù)據(jù)集時(shí)候進(jìn)行了變換transforms。 加入對(duì)GPU的支持，首先判斷torch.cuda.is_available(),然后決定使用GPU or CPU
import torch
import torch.nn as nn
import torch.optim as optim
from torch.optim import lr_scheduler
from torch.utils.data import DataLoader
import torchvision
from torchvision.transforms import transforms
from torchvision import models
from torchvision.models import ResNet
import numpy as np
import matplotlib.pyplot as plt
import os
import utils
data_dir = './data/hymenoptera_data'
train_dataset = torchvision.datasets.ImageFolder(root=os.path.join(data_dir, 'train'),
transform=transforms.Compose(
[
transforms.RandomResizedCrop(224),
transforms.RandomHorizontalFlip(),
transforms.ToTensor(),
transforms.Normalize(
mean=(0.485, 0.456, 0.406),
std=(0.229, 0.224, 0.225))
]))
val_dataset = torchvision.datasets.ImageFolder(root=os.path.join(data_dir, 'val'),
transform=transforms.Compose(
[
transforms.RandomResizedCrop(224),
transforms.RandomHorizontalFlip(),
transforms.ToTensor(),
transforms.Normalize(
mean=(0.485, 0.456, 0.406),
std=(0.229, 0.224, 0.225))
]))
train_dataloader = DataLoader(dataset=train_dataset, batch_size=4, shuffle=4)
val_dataloader = DataLoader(dataset=val_dataset, batch_size=4, shuffle=4)
# 類別名稱
class_names = train_dataset.classes
print('class_names:{}'.format(class_names))
# 訓(xùn)練設(shè)備 CPU/GPU
device = torch.device('cuda:0' if torch.cuda.is_available() else 'cpu')
print('trian_device:{}'.format(device.type))
# 隨機(jī)顯示一個(gè)batch
plt.figure()
utils.imshow(next(iter(train_dataloader)))
plt.show()
獲取預(yù)訓(xùn)練模型
torchvision.models
torchvision中包含了一些常見的預(yù)訓(xùn)練模型：
AlexNet, VGG, SqueezeNet, Resnet奋姿，Inception, DenseNet
此次實(shí)驗(yàn)采用ResNet18網(wǎng)絡(luò)模型锄开。
在torchvision.models中包含resnet18，首先會(huì)實(shí)例化一個(gè)ResNet網(wǎng)絡(luò)称诗， 然后model.load_dict()加載預(yù)訓(xùn)練好的模型萍悴。
def resnet18(pretrained=False, **kwargs):
"""Constructs a ResNet-18 model.
Args:
pretrained (bool): If True, returns a model pre-trained on ImageNet
"""
model = ResNet(BasicBlock, [2, 2, 2, 2], **kwargs)
if pretrained:
model.load_state_dict(model_zoo.load_url(model_urls['resnet18']))
return model
torchvision 默認(rèn)將模型保存在/home/.torch/models路徑。
預(yù)訓(xùn)練模型文件：
-
代碼
加載預(yù)訓(xùn)練模型寓免。需要注意的地方：修改ResNet最后一個(gè)全連接層的輸出個(gè)數(shù)癣诱，二分類問(wèn)題需要將輸出個(gè)數(shù)改為2。
# -------------------------模型選擇再榄，優(yōu)化方法狡刘， 學(xué)習(xí)率策略----------------------
model = models.resnet18(pretrained=True)
# 全連接層的輸入通道in_channels個(gè)數(shù)
num_fc_in = model.fc.in_features
# 改變?nèi)B接層，2分類問(wèn)題困鸥，out_features = 2
model.fc = nn.Linear(num_fc_in, 2)
# 模型遷移到CPU/GPU
model = model.to(device)
# 定義損失函數(shù)
loss_fc = nn.CrossEntropyLoss()
# 選擇優(yōu)化方法
optimizer = optim.SGD(model.parameters(), lr=0.0001, momentum=0.9)
# 學(xué)習(xí)率調(diào)整策略
# 每7個(gè)epoch調(diào)整一次
exp_lr_scheduler = lr_scheduler.StepLR(optimizer=optimizer, step_size=10, gamma=0.5) # step_size
訓(xùn)練嗅蔬，測(cè)試網(wǎng)絡(luò)
Epoch: 訓(xùn)練50個(gè)epoch
注意地方： 訓(xùn)練時(shí)候，需要調(diào)用model.train()將模型設(shè)置為訓(xùn)練模式疾就。測(cè)試時(shí)候澜术，調(diào)用model.eval() 將模型設(shè)置為測(cè)試模型，否則訓(xùn)練和測(cè)試結(jié)果不正確猬腰。
# ----------------訓(xùn)練過(guò)程-----------------
num_epochs = 50
for epoch in range(num_epochs):
running_loss = 0.0
exp_lr_scheduler.step()
for i, sample_batch in enumerate(train_dataloader):
inputs = sample_batch[0]
labels = sample_batch[1]
model.train()
# GPU/CPU
inputs = inputs.to(device)
labels = labels.to(device)
optimizer.zero_grad()
# foward
outputs = model(inputs)
# loss
loss = loss_fc(outputs, labels)
# loss求導(dǎo)鸟废，反向
loss.backward()
# 優(yōu)化
optimizer.step()
#
running_loss += loss.item()
# 測(cè)試
if i % 20 == 19:
correct = 0
total = 0
model.eval()
for images_test, labels_test in val_dataloader:
images_test = images_test.to(device)
labels_test = labels_test.to(device)
outputs_test = model(images_test)
_, prediction = torch.max(outputs_test, 1)
correct += (torch.sum((prediction == labels_test))).item()
# print(prediction, labels_test, correct)
total += labels_test.size(0)
print('[{}, {}] running_loss = {:.5f} accurcay = {:.5f}'.format(epoch + 1, i + 1, running_loss / 20,
correct / total))
running_loss = 0.0
# if i % 10 == 9:
# print('[{}, {}] loss={:.5f}'.format(epoch+1, i+1, running_loss / 10))
# running_loss = 0.0
print('training finish !')
torch.save(model.state_dict(), './model/model_2.pth')
訓(xùn)練輸出結(jié)果
隨著訓(xùn)練次數(shù)增加，accuracy基本上是上升趨勢(shì)姑荷，最終達(dá)到93%的準(zhǔn)確率盒延。
完整代碼
import torch
import torch.nn as nn
import torch.optim as optim
from torch.optim import lr_scheduler
from torch.utils.data import DataLoader
import torchvision
from torchvision.transforms import transforms
from torchvision import models
from torchvision.models import ResNet
import numpy as np
import matplotlib.pyplot as plt
import os
import utils
data_dir = './data/hymenoptera_data'
train_dataset = torchvision.datasets.ImageFolder(root=os.path.join(data_dir, 'train'),
transform=transforms.Compose(
[
transforms.RandomResizedCrop(224),
transforms.RandomHorizontalFlip(),
transforms.ToTensor(),
transforms.Normalize(
mean=(0.485, 0.456, 0.406),
std=(0.229, 0.224, 0.225))
]))
val_dataset = torchvision.datasets.ImageFolder(root=os.path.join(data_dir, 'val'),
transform=transforms.Compose(
[
transforms.RandomResizedCrop(224),
transforms.RandomHorizontalFlip(),
transforms.ToTensor(),
transforms.Normalize(
mean=(0.485, 0.456, 0.406),
std=(0.229, 0.224, 0.225))
]))
train_dataloader = DataLoader(dataset=train_dataset, batch_size=4, shuffle=4)
val_dataloader = DataLoader(dataset=val_dataset, batch_size=4, shuffle=4)
# 類別名稱
class_names = train_dataset.classes
print('class_names:{}'.format(class_names))
# 訓(xùn)練設(shè)備 CPU/GPU
device = torch.device('cuda:0' if torch.cuda.is_available() else 'cpu')
print('trian_device:{}'.format(device.type))
# 隨機(jī)顯示一個(gè)batch
#plt.figure()
#utils.imshow(next(iter(train_dataloader)))
#plt.show()
# -------------------------模型選擇，優(yōu)化方法鼠冕， 學(xué)習(xí)率策略----------------------
model = models.resnet18(pretrained=True)
# 全連接層的輸入通道in_channels個(gè)數(shù)
num_fc_in = model.fc.in_features
# 改變?nèi)B接層添寺，2分類問(wèn)題，out_features = 2
model.fc = nn.Linear(num_fc_in, 2)
# 模型遷移到CPU/GPU
model = model.to(device)
# 定義損失函數(shù)
loss_fc = nn.CrossEntropyLoss()
# 選擇優(yōu)化方法
optimizer = optim.SGD(model.parameters(), lr=0.0001, momentum=0.9)
# 學(xué)習(xí)率調(diào)整策略
# 每7個(gè)epoch調(diào)整一次
exp_lr_scheduler = lr_scheduler.StepLR(optimizer=optimizer, step_size=10, gamma=0.5) # step_size
# ----------------訓(xùn)練過(guò)程-----------------
num_epochs = 50
for epoch in range(num_epochs):
running_loss = 0.0
exp_lr_scheduler.step()
for i, sample_batch in enumerate(train_dataloader):
inputs = sample_batch[0]
labels = sample_batch[1]
model.train()
# GPU/CPU
inputs = inputs.to(device)
labels = labels.to(device)
optimizer.zero_grad()
# foward
outputs = model(inputs)
# loss
loss = loss_fc(outputs, labels)
# loss求導(dǎo)懈费，反向
loss.backward()
# 優(yōu)化
optimizer.step()
#
running_loss += loss.item()
# 測(cè)試
if i % 20 == 19:
correct = 0
total = 0
model.eval()
for images_test, labels_test in val_dataloader:
images_test = images_test.to(device)
labels_test = labels_test.to(device)
outputs_test = model(images_test)
_, prediction = torch.max(outputs_test, 1)
correct += (torch.sum((prediction == labels_test))).item()
# print(prediction, labels_test, correct)
total += labels_test.size(0)
print('[{}, {}] running_loss = {:.5f} accurcay = {:.5f}'.format(epoch + 1, i + 1, running_loss / 20,
correct / total))
running_loss = 0.0
# if i % 10 == 9:
# print('[{}, {}] loss={:.5f}'.format(epoch+1, i+1, running_loss / 10))
# running_loss = 0.0
print('training finish !')
torch.save(model.state_dict(), './model/model_2.pth')
End
參考：
https://pytorch.org/tutorials/beginner/transfer_learning_tutorial.html
https://blog.csdn.net/sunqiande88/article/details/80100891
