【3】迁移学习模型
前言
主要简述一下训练代码
一、安装相关模块
#xingyun的笔记本
print('============================xingyun的笔记本=============================')
%pip install d2l
%pip install Ipython
%pip install efficientnet_pytorch #（可选）
%pip install timm
二、训练代码
整段代码大致分为四块： 管理预训练模型、模型训练、可视化结果、类别函数调用。
2.1. 管理预训练模型
用于管理要使用的迁移学习模型（可添加），这里主要是对EfficientNet系列模型 、ResNet系列模型、MobileNet系列模型进行迁移学习。
import collections
import math
import os
import shutil
import pandas as pd
import torch
import torchvision
import timm
from torch import nn
from d2l import torch as d2l
import re
from itertools import product
from sklearn.metrics import classification_report, confusion_matrix, roc_curve, auc
import seaborn as sns
import numpy as np
import matplotlib.pyplot as plt
import matplotlib.colors as mcolors
from efficientnet_pytorch import EfficientNet
from sklearn.model_selection import KFold
from torchvision.models import resnet101, resnet152, resnet18, resnet34, resnet50,mobilenet,mobilenet_v2,mobilenet_v3_large, mobilenet_v3_small, mobilenetv2, mobilenetv3
class FineTuneModel:
"""
管理预训练模型：
1. EfficientNet系列模型
3. ResNet系列模型
4. MobileNet系列模型
"""
def __init__(self, devices, num_classes,model_name):
self.devices = devices
self.num_classes = num_classes
self.model_name = model_name
def get_efficientnet(self):
"""
微调EfficientNet模型。
:param model_name: EfficientNet模型的版本，如'efficientnet-b0'到'efficientnet-b7'。
:return: 微调后的模型。
"""
# 加载预训练的EfficientNet模型
finetune_net = EfficientNet.from_pretrained(self.model_name)
# 替换输出层
num_ftrs = finetune_net._fc.in_features
finetune_net._fc = nn.Linear(num_ftrs, self.num_classes)
# 将模型参数分配到设备上
finetune_net = finetune_net.to(self.devices[0])
# 冻结所有层（除了最后的全连接层）
for name, param in finetune_net.named_parameters():
if 'fc' not in name: # 不冻结全连接层
param.requires_grad = False
# 确保全连接层的参数可训练
for param in finetune_net._fc.parameters():
param.requires_grad = True
return finetune_net
def get_mobilenet(self):
"""
加载预训练的MobileNet模型并进行微调设置。
'mobilenet',
'mobilenet_v2',
'mobilenet_v3_large',
'mobilenet_v3_small',
'mobilenetv2',
'mobilenetv3',
"""
# 加载预训练的MobileNetV2模型
base_model_func = getattr(torchvision.models, self.model_name)
base_model = base_model_func(pretrained=True)
if self.model_name == "mobilenet_v2":
num_features = base_model.classifier[-1].in_features
# 定义一个新的分类头
classifier = nn.Sequential(
nn.Linear(num_features, 256),
nn.ReLU(),
nn.Linear(256, self.num_classes)
)
# 替换原有的分类头
base_model.classifier = classifier
else:
# 获取最后一个卷积层的输出通道数量
num_features = base_model.features[-1].out_channels
# 定义一个新的分类头
classifier = nn.Sequential(
nn.Linear(num_features, 256),
nn.ReLU(),
nn.Linear(256, self.num_classes)
)
# 替换原有的分类头
base_model.classifier = classifier
# 将模型参数分配到指定设备
base_model = base_model.to(self.devices[0])
# 冻结特征提取器的参数
for name, param in base_model.named_parameters():
if 'classifier' not in name: # 确保只冻结特征提取部分的参数
param.requires_grad = False
return base_model
def get_resnet(self):
#加载预训练的resnet模型
"""
resnet101,
resnet152,
resnet18,
resnet34,
resnet50
"""
# 从torchvision.models模块中动态获取模型
base_model_func = getattr(torchvision.models, self.model_name)
base_model = base_model_func(pretrained=True)
num_features = base_model.fc.in_features
# 定义一个新的分类头
classifier = nn.Sequential(
nn.Linear(num_features, 256),
nn.ReLU(),
nn.Linear(256,self.num_classes)
)
# 替换原有的全连接层（分类头）
base_model.fc = classifier
# 将模型参数分配到指定设备
base_model = base_model.to(self.devices[0])
# 冻结特征提取器的参数
for name, param in base_model.named_parameters():
if 'fc' not in name: # 确保只冻结特征提取部分的参数
param.requires_grad = False
return base_model
2.2. 模型训练代码
包括数据处理、模型训练、参数调优、模型保存等。
class MyImageClassifier:
"""
1. 数据处理
2. 模型训练
3. 参数调优
4. 模型保存
"""
def __init__(self, data_dir, target_dir, batch_size, valid_ratio,train_folder,test_folder):
self.data_dir = data_dir
self.target_dir = target_dir
self.batch_size = batch_size
self.valid_ratio = valid_ratio
self.train_folder = train_folder
self.test_folder = test_folder
def read_csv_labels(self, fname):
"""读取fname来给标签字典返回一个文件名"""
with open(fname, 'r') as f:
# 跳过文件头行(列名)
lines = f.readlines()[1:]
tokens = [l.rstrip().split(',') for l in lines]
return dict(((name, label) for name, label in tokens))
def copyfile(self,filename, target_dir):
"""将文件复制到目标目录"""
os.makedirs(target_dir, exist_ok=True)
shutil.copy(filename, target_dir)
def reorg_train_valid(self,labels):
"""将验证集从原始的训练集中拆分出来"""
# 训练数据集中样本最少的类别中的样本数
n = collections.Counter(labels.values()).most_common()[-1][1]
# 验证集中每个类别的样本数
n_valid_per_label = max(1, math.floor(n * self.valid_ratio))
label_count = {}
for train_file in os.listdir(os.path.join(self.data_dir, self.train_folder)):
label = labels[train_file.split('.')[0]]
fname = os.path.join(self.data_dir, self.train_folder, train_file)
self.copyfile(fname, os.path.join(self.target_dir, 'train_valid_test',
'train_valid', label))
if label not in label_count or label_count[label] < n_valid_per_label:
self.copyfile(fname, os.path.join(self.target_dir, 'train_valid_test',
'valid', label))
label_count[label] = label_count.get(label, 0) + 1
else:
self.copyfile(fname, os.path.join(self.target_dir, 'train_valid_test',
'train', label))
return n_valid_per_label
def reorg_test(self):
"""在预测期间整理测试集，以方便读取"""
for test_file in os.listdir(os.path.join(self.data_dir, self.test_folder)):
self.copyfile(os.path.join(self.data_dir, self.test_folder, test_file),
os.path.join(self.target_dir, 'train_valid_test', 'test',
'unknown'))
def reorg_san_data(self,labels_csv):
labels = self.read_csv_labels(os.path.join(self.data_dir,labels_csv))
self.reorg_train_valid(labels)
self.reorg_test()
print('# 训练样本 :', len(labels))
print('# 类别 :', len(set(labels.values())))
"""
以上为数据整理函数
"""
def classes(self):
class_to_idx = {}
# 遍历数据集文件夹中的子文件夹（每个子文件夹代表一个类别）
for idx, class_name in enumerate(sorted(os.listdir(os.path.join(self.target_dir, 'train_valid_test', 'valid')))):
if class_name.startswith('.'):
continue
class_dir = os.path.join(os.path.join(self.target_dir, 'train_valid_test', 'valid'), class_name) # 类别文件夹路径
if os.path.isdir(class_dir):
class_to_idx[idx] = class_name
print(class_to_idx)
print("============================")
return class_to_idx
#统计划分的训练集、验证集数量
def count_samples(self):
"""统计每个类别训练集和验证集的数量"""
train_valid_test_dirs = ['train', 'valid']
data_counts = {'class': []}
for dir_name in train_valid_test_dirs:
class_dir = os.path.join(self.target_dir, 'train_valid_test', dir_name)
if dir_name not in data_counts:
data_counts[dir_name] = []
for class_name in os.listdir(class_dir):
if class_name.startswith('.'):
continue
class_sub_dir = os.path.join(class_dir, class_name)
if os.path.isdir(class_sub_dir):
if class_name not in data_counts['class']:
data_counts['class'].append(class_name)
for key in train_valid_test_dirs:
if key not in data_counts:
data_counts[key] = [0] * len(data_counts['class'])
else:
data_counts[key].append(0)
data_counts[dir_name][data_counts['class'].index(class_name)] += len(os.listdir(class_sub_dir))
df = pd.DataFrame(data_counts)
return df
def shuju_zq_jz(self,batch_size):
#数据增强
transform_train = torchvision.transforms.Compose([
# 随机裁剪图像，所得图像为原始面积的0.08～1之间，高宽比在3/4和4/3之间。
# 然后，缩放图像以创建224x224的新图像
torchvision.transforms.RandomResizedCrop(224, scale=(0.08, 1.0),
ratio=(3.0/4.0, 4.0/3.0)),
torchvision.transforms.RandomHorizontalFlip(),
# 随机更改亮度，对比度和饱和度
torchvision.transforms.ColorJitter(brightness=0.4,
contrast=0.4,
saturation=0.4),
#转换为张量格式
torchvision.transforms.ToTensor(),
# 标准化图像的每个通道
torchvision.transforms.Normalize([0.485, 0.456, 0.406],
[0.229, 0.224, 0.225])])
#测试时，我们只使用确定性的图像预处理操作
transform_test = torchvision.transforms.Compose([
torchvision.transforms.Resize(256),
# 从图像中心裁切224x224大小的图片
torchvision.transforms.CenterCrop(224),
torchvision.transforms.ToTensor(),
torchvision.transforms.Normalize([0.485, 0.456, 0.406],
[0.229, 0.224, 0.225])])
#读取整理后的含原始图像文件的数据集
train_ds, train_valid_ds = [torchvision.datasets.ImageFolder(
os.path.join(self.target_dir, 'train_valid_test', folder),
transform=transform_train) for folder in ['train', 'train_valid']]
valid_ds, test_ds = [torchvision.datasets.ImageFolder(
os.path.join(self.target_dir, 'train_valid_test', folder),
transform=transform_test) for folder in ['valid', 'test']]
train_iter, train_valid_iter = [torch.utils.data.DataLoader(
dataset, batch_size, shuffle=True, drop_last=True)
for dataset in (train_ds, train_valid_ds)]
valid_iter = torch.utils.data.DataLoader(valid_ds, batch_size, shuffle=False,
drop_last=False)
test_iter = torch.utils.data.DataLoader(test_ds, batch_size, shuffle=False,
drop_last=False)
return train_iter,valid_iter,test_iter,train_valid_iter
"""
以上为数据处理函数
"""
#微调预训练模型
def get_net(self,devices,num_classes,model_name,model_leibie):
fine_tune_model = FineTuneModel(d2l.try_all_gpus(), num_classes=num_classes,model_name = model_name) # 使用微调模型
if model_leibie == 'get_efficientnet':
base_model = fine_tune_model.get_efficientnet()
elif model_leibie == 'get_mobilenet':
base_model = fine_tune_model.get_mobilenet()
elif model_leibie == 'get_resnet':
base_model = fine_tune_model.get_resnet()
return base_model
def evaluate_loss(self,data_iter, net, devices):
loss = nn.CrossEntropyLoss(reduction='none') #reduction='none'表示不对损失进行平均或求和，而是返回每个样本的损失值。
l_sum, n = 0.0, 0
for features, labels in data_iter:
features, labels = features.to(devices[0]), labels.to(devices[0])
outputs = net(features)
l = loss(outputs, labels)
l_sum += l.sum()
n += labels.numel() #累加样本数量到n中，labels.numel()返回标签张量中元素的个数
return (l_sum / n).to('cpu') #计算所有样本的平均损失值，并将其移动到CPU上返回
def train(self, net, train_iter, valid_iter, num_epochs, lr, wd, devices, lr_period, lr_decay):
net = nn.DataParallel(net, device_ids=devices).to(devices[0])
trainer = torch.optim.SGD((param for param in net.parameters() if param.requires_grad), lr=lr, momentum=0.9, weight_decay=wd)
scheduler = torch.optim.lr_scheduler.StepLR(trainer, lr_period, lr_decay)
num_batches, timer = len(train_iter), d2l.Timer()
legend = ['train loss', 'train acc', 'valid loss', 'valid acc'] # Add valid loss and acc
animator = d2l.Animator(xlabel='epoch', xlim=[1, num_epochs], legend=legend)
loss = nn.CrossEntropyLoss(reduction='none')
best_acc = 0
best_model_path = ""
measures_list = []
examples_sec_list = []
for epoch in range(num_epochs):
metric = d2l.Accumulator(3)
net.train() # Switch to training mode
for i, (features, labels) in enumerate(train_iter):
timer.start()
features, labels = features.to(devices[0]), labels.to(devices[0])
trainer.zero_grad()
output = net(features)
l = loss(output, labels).sum()
l.backward()
trainer.step()
metric.add(l, labels.shape[0], d2l.accuracy(output, labels))
timer.stop()
if (i + 1) % (num_batches // 5) == 0 or i == num_batches - 1:
animator.add(epoch + (i + 1) / num_batches, (metric[0] / metric[1], metric[2] / metric[1], None, None))
measures = f'train loss {metric[0] / metric[1]:.3f}, train acc {metric[2] / metric[1]:.3f}'
if valid_iter is not None:
net.eval() # Switch to evaluation mode
valid_metric = d2l.Accumulator(3)
with torch.no_grad():
for valid_features, valid_labels in valid_iter:
valid_features, valid_labels = valid_features.to(devices[0]), valid_labels.to(devices[0])
valid_output = net(valid_features)
valid_l = loss(valid_output, valid_labels).sum()
valid_metric.add(valid_l, valid_labels.shape[0], d2l.accuracy(valid_output, valid_labels))
valid_acc = valid_metric[2] / valid_metric[1]
animator.add(epoch + 1, (None, None, valid_metric[0] / valid_metric[1], valid_acc))
measures += f', valid loss {valid_metric[0] / valid_metric[1]:.3f}, valid acc {valid_acc:.3f}'
if valid_acc > best_acc:
best_acc = valid_acc
best_model_path = f'model_bests.pth'
torch.save(net, best_model_path)
print(f"Best model saved to {best_model_path} with accuracy {best_acc:.3f}")
measures_list.append(measures)
examples_sec = f'epoch {epoch}, {metric[1] * num_epochs / timer.sum():.1f} examples/sec on {str(devices)}'
examples_sec_list.append(examples_sec)
print(f'epoch {epoch}, ' + measures + f'\n{metric[1] * num_epochs / timer.sum():.1f} examples/sec on {str(devices)}')
scheduler.step()
for i, measure in enumerate(measures_list):
print(f"Epoch {(i+1)}: {measure}")
print(examples_sec_list)
return measures_list, examples_sec_list
def get_valid_acc(self,measures):
return float(re.search(r'valid acc (\d+\.\d+)', measures).group(1))
# 训练参数调优
def train_parameter_tuning(self, param_grid,num_classes,batch_size,model_name,model_leibie):
# 使用网格搜索来搜索最佳超参数组合
best_accuracy = 0
best_params = None
# 初始化列表，用于存储包含验证准确率和参数的元组
acc_param_list = []
measures_lt = []
for params in product(*param_grid.values()):
param_dict = dict(zip(param_grid.keys(), params))
print("Training with params:", param_dict)
# 创建和训练模型
net = self.get_net(d2l.try_all_gpus(),num_classes,model_name,model_leibie)
train_iter,valid_iter = self.shuju_zq_jz(batch_size)[0],self.shuju_zq_jz(batch_size)[1]
measures_list,examples_sec_list = self.train(net, train_iter, valid_iter, **param_dict,devices = d2l.try_all_gpus())
# 在验证集上评估模型性能
# 使用正则表达式提取valid acc对应的数值
best_measures = max(measures_list, key=self.get_valid_acc)
valid_acc = float(re.search(r'valid acc (\d+\.\d+)', best_measures).group(1))
print(best_measures)
# 将验证准确率和参数字典合并为一个元组，并添加到列表中
acc_param_list.append((valid_acc, param_dict))
measures_lt.append(best_measures)
# net.load_state_dict(torch.load("model_best.pth")) # 加载最佳模型
net = torch.load('model_bests.pth')
if valid_acc > best_accuracy:
best_accuracy = valid_acc
best_params = param_dict
best_net = net # 这里的最佳网络是从最佳模型加载的
for i,measure in enumerate(measures_lt):
print(f"Trial {i+1}:")
print(measure)
print("========================================================")
print()
best_acc_index = max(range(len(acc_param_list)), key=lambda i: acc_param_list[i][0])
best_accuracy = acc_param_list[best_acc_index][0]
best_params = acc_param_list[best_acc_index][1]
print("================================================")
print("Best accuracy:", best_accuracy)
print("Best params:", best_params)
print()
for i, (acc, params) in enumerate(acc_param_list):
print(f"Trial {i+1}: valid acc {acc}, params {params}")
return best_net # 这是从最佳模型加载的网络
"""
以上为模型训练以及参数调优函数
"""
#保存训练得到的模型权重文件
def save_model(self,model_path,model_path_zheng,best_net):
torch.save(best_net.state_dict(), model_path) #只保存模型的参数
torch.save(best_net, model_path_zheng) #保存整个模型
print(f"Model saved to {model_path}")
2.3. 可视化结果
包括查看模型在验证集上的每一类的准确率、分类报告、混淆矩阵、AUC-ROC曲线
class ViewResult:
"""
查看训练效果:
1. 查看每一类在验证集上的准确率
2. 查看precision，recall和f1-score(即分类报告)
3. 查看混淆矩阵
4. 查看AUC-ROC曲线
"""
def __init__(self, best_net, valid_iter, devices, classes):
self.net = best_net
self.valid_iter = valid_iter
self.devices = devices
self.classes = classes
self.num_classes = len(classes)
def view_result(self):
print(self.num_classes)
class_correct = [0.] * self.num_classes
class_total = [0.] * self.num_classes
y_test, y_pred = [], []
X_test = []
with torch.no_grad():
for images, labels in self.valid_iter:
X_test.extend([_ for _ in images])
outputs = self.net(images.to(self.devices[0]))
_, predicted = torch.max(outputs, 1)
predicted = predicted.cpu()
c = (predicted == labels).squeeze()
for i, label in enumerate(labels):
class_correct[label] += c[i].item()
class_total[label] += 1
y_pred.extend(predicted.numpy())
y_test.extend(labels.cpu().numpy())
for i in range(self.num_classes):
if class_total[i] != 0:
accuracy = 100 * class_correct[i] / class_total[i]
else:
accuracy = 0
print(f"Accuracy of {self.classes[i]:5s}: {accuracy:2.0f}%")
#分类报告
try:
cr = classification_report(y_test, y_pred, target_names=list(self.classes.values()))
print(cr)
except Exception as e:
print("An error occurred while generating the classification report:", str(e))
#混淆矩阵
cm = confusion_matrix(y_test, y_pred)
labels = pd.DataFrame(cm).applymap(lambda v: f"{v}" if v != 0 else f"")
plt.figure(figsize=(25, 20))
sns.heatmap(cm, annot=labels, fmt='s', xticklabels=self.classes.items(), yticklabels=self.classes.items(), linewidths=0.1)
plt.show()
def evaluate_roc(self, num_classes):
self.net.eval()
y_true = []
y_score = []
for X, y in self.valid_iter:
X, y = X.to(self.devices[0]), y.to(self.devices[0])
y_true.append(y.cpu().numpy())
y_score.append(self.net(X).detach().cpu().numpy())
y_true = np.concatenate(y_true)
y_score = np.concatenate(y_score)
fpr = dict()
tpr = dict()
roc_auc = dict()
for i in range(num_classes):
y_true_i = np.where(y_true == i, 1, 0)
y_score_i = y_score[:, i]
fpr[i], tpr[i], _ = roc_curve(y_true_i, y_score_i)
roc_auc[i] = auc(fpr[i], tpr[i])
plt.figure(figsize=(15, 15))
colors = list(mcolors.CSS4_COLORS.values())
colors = colors[:num_classes]
for i in range(num_classes):
plt.plot(fpr[i], tpr[i], color=colors[i], lw=2, label=f'Class {i}, AUC = {roc_auc[i]:.2f}')
plt.plot([0, 1], [0, 1], color='navy', lw=2, linestyle='--')
plt.xlim([0.0, 1.0])
plt.ylim([0.0, 1.05])
plt.xlabel('False Positive Rate')
plt.ylabel('True Positive Rate')
plt.title('Receiver Operating Characteristic for Multi-class Classification')
plt.legend(loc="lower right")
plt.show()
2.4. 类别函数
调用之前所定义的类，实现不同分类类别的模型的训练
def leibie_class(leibie,num_classes,batch_size,valid_ratio,param_grid,model_path,target_dir,model_name,model_leibie,train_folder,test_folder,model_path_zheng):
"""
leibie: #填写是几分类的标签文件，如要进行十分类，则填写labels.csv;要进行五分类，则填写labels_5.csv；要进行二分类，填写labels_2.csv.
num_classes : #填写与leibie对应的类别数字，如要进行十分类，填写数字10，以此类推。
batch_size : #批量大小，即每次处理的样本数量。
valid_ratio : #验证集所占比例，如为0.3，则代表验证集占比为30%；即验证集：训练集==3：7。以此类推
param_grid : #定义要调整的参数范围,通过网格搜索，遍历出最佳模型，获取对应的参数。
model_path : #最佳模型权重文件保存路径，如'/kaggle/working/model_xcy_shi_1.pth'，可以更改'/model_xcy_shi_1.pth',前面的‘/kaggle/working/’不能改变。
target_dir : #整理后目标文件存放的目录，如‘/kaggle/working/my_directory_shi’，代表是按十分类整理的文件；‘/kaggle/working/my_directory_wu’则代表是按五分类整理的文件。以此类推。
model_name : #加载模型的名字。例如：'resnet34'
model_leibie : #加载模型所属类别函数。例如： 'get_resnet'
train_folder : #原始训练集存放文件夹
test_folder : #原始测试集存放文件夹
model_path_zheng : #完整模型存放路径
"""
data_dir = "/kaggle/input/sanyeqing/" #存放原始数据的目录
image_classifier = MyImageClassifier(data_dir, target_dir, batch_size, valid_ratio,train_folder,test_folder) #图像分类训练模型类的实例化
#调用类中的函数：
image_classifier.reorg_san_data(leibie) #十分类(labels.csv为十分类，labels_5.csv为五分类......)
valid_iter=image_classifier.shuju_zq_jz(batch_size)[1] #数据增强处理函数,
class_to_idx=image_classifier.classes() #返回分类标签与索引对应函数
print(class_to_idx)
best_net = image_classifier.train_parameter_tuning(param_grid,num_classes,batch_size,model_name,model_leibie) #十分类(10为十分类,5则为五分类......)
# print("===================================================")
# print("最终的保存模型：")
# image_classifier.save_model(model_path,model_path_zheng,best_net)
# print();print()
#训练结果可视化：
# result_viewer = ViewResult(best_net, valid_iter, devices=d2l.try_all_gpus(), classes=class_to_idx)
# result_viewer.view_result()
# result_viewer.evaluate_roc(num_classes)
# 使用示例
if __name__ == '__main__':
# 定义要调整的参数范围
param_grid = {
'num_epochs': [201],
'lr': [1e-4],
'wd': [1e-4],
'lr_period': [2],
'lr_decay': [1]
}
print("这是省份分类：")
leibie_class("labels_hun_finally_2.csv",
num_classes=3,
batch_size=128,
valid_ratio=0.2,
param_grid=param_grid,
# model_path='/kaggle/working/model_wht_wu.pth',
model_path=None,
target_dir='/kaggle/working/my_directory_er',
model_name='mobilenet_v3_large',
model_leibie='get_mobilenet',
train_folder = 'train_hun_finally',
test_folder='test_hun_finally',
model_path_zheng=None) #省份分类
总结
主要是运用迁移学习的方法，将预训练模型在自定义的数据集上进行训练。
2024/6/12
