新网站如何做百度收录,建设银行网站特点分析,网件r6300v2,浏览器打不开二级网页【3】迁移学习模型 文章目录 前言一、安装相关模块二、训练代码2.1. 管理预训练模型2.2. 模型训练代码2.3. 可视化结果2.4. 类别函数 总结 前言
主要简述一下训练代码
三叶青图像识别研究简概 一、安装相关模块
#xingyun的笔记本
print(xingyun的笔记本)
%pip install d2l
%… 【3】迁移学习模型 文章目录 前言一、安装相关模块二、训练代码2.1. 管理预训练模型2.2. 模型训练代码2.3. 可视化结果2.4. 类别函数 总结 前言
主要简述一下训练代码
三叶青图像识别研究简概 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, mobilenetv3class 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 devicesself.num_classes num_classesself.model_name model_namedef 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_featuresfinetune_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 Truereturn finetune_netdef 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(pretrainedTrue)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 classifierelse:# 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 Falsereturn base_modeldef 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(pretrainedTrue)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 Falsereturn 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_dirself.target_dir target_dirself.batch_size batch_sizeself.valid_ratio valid_ratioself.train_folder train_folderself.test_folder test_folderdef 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_okTrue)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) 1else:self.copyfile(fname, os.path.join(self.target_dir, train_valid_test,train, label))return n_valid_per_labeldef 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(.):continueclass_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_nameprint(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(.):continueclass_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 dfdef 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0.081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(brightness0.4,contrast0.4,saturation0.4),#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),transformtransform_train) for folder in [train, train_valid]]valid_ds, test_ds [torchvision.datasets.ImageFolder(os.path.join(self.target_dir, train_valid_test, folder),transformtransform_test) for folder in [valid, test]]train_iter, train_valid_iter [torch.utils.data.DataLoader(dataset, batch_size, shuffleTrue, drop_lastTrue)for dataset in (train_ds, train_valid_ds)]valid_iter torch.utils.data.DataLoader(valid_ds, batch_size, shuffleFalse,drop_lastFalse)test_iter torch.utils.data.DataLoader(test_ds, batch_size, shuffleFalse,drop_lastFalse)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_classesnum_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_modeldef evaluate_loss(self,data_iter, net, devices):loss nn.CrossEntropyLoss(reductionnone) #reductionnone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, 0for 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_idsdevices).to(devices[0])trainer torch.optim.SGD((param for param in net.parameters() if param.requires_grad), lrlr, momentum0.9, weight_decaywd)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 accanimator d2l.Animator(xlabelepoch, xlim[1, num_epochs], legendlegend)loss nn.CrossEntropyLoss(reductionnone)best_acc 0best_model_path measures_list []examples_sec_list []for epoch in range(num_epochs):metric d2l.Accumulator(3)net.train() # Switch to training modefor 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 ftrain loss {metric[0] / metric[1]:.3f}, train acc {metric[2] / metric[1]:.3f}if valid_iter is not None:net.eval() # Switch to evaluation modevalid_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_accbest_model_path fmodel_bests.pthtorch.save(net, best_model_path)print(fBest model saved to {best_model_path} with accuracy {best_acc:.3f})measures_list.append(measures)examples_sec fepoch {epoch}, {metric[1] * num_epochs / timer.sum():.1f} examples/sec on {str(devices)}examples_sec_list.append(examples_sec)print(fepoch {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(fEpoch {(i1)}: {measure})print(examples_sec_list)return measures_list, examples_sec_listdef get_valid_acc(self,measures):return float(re.search(rvalid 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 0best_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, keyself.get_valid_acc)valid_acc float(re.search(rvalid 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_accbest_params param_dictbest_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(fTrial {i1}:)print(measure)print()print()best_acc_index max(range(len(acc_param_list)), keylambda 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(fTrial {i1}: 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(fModel 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precisionrecall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_netself.valid_iter valid_iterself.devices devicesself.classes classesself.num_classes len(classes)def view_result(self):print(self.num_classes)class_correct [0.] * self.num_classesclass_total [0.] * self.num_classesy_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] 1y_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 0print(fAccuracy of {self.classes[i]:5s}: {accuracy:2.0f}%)#分类报告try:cr classification_report(y_test, y_pred, target_nameslist(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, annotlabels, fmts, xticklabelsself.classes.items(), yticklabelsself.classes.items(), linewidths0.1)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], colorcolors[i], lw2, labelfClass {i}, AUC {roc_auc[i]:.2f})plt.plot([0, 1], [0, 1], colornavy, lw2, 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(loclower 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37。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resnet34model_leibie : #加载模型所属类别函数。例如 get_resnettrain_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_iterimage_classifier.shuju_zq_jz(batch_size)[1] #数据增强处理函数,class_to_idximage_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, devicesd2l.try_all_gpus(), classesclass_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_classes3,batch_size128,valid_ratio0.2,param_gridparam_grid,
# model_path/kaggle/working/model_wht_wu.pth,model_pathNone,target_dir/kaggle/working/my_directory_er,model_namemobilenet_v3_large,model_leibieget_mobilenet,train_folder train_hun_finally,test_foldertest_hun_finally,model_path_zhengNone) #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
