分类任务实现模型（投票式）集成代码模版

简介

本实验使用上一博客的深度学习分类模型训练代码模板-CSDN博客，自定义投票式集成，手动实现模型集成（投票法）的代码。最后通过tensorboard进行可视化，对每个基学习器的性能进行对比，直观的看出模型集成的作用。

代码

# -*- coding:utf-8 -*-
import os
import torch
import torchvision
import torchmetrics
import torch.nn as nn
import my_utils as utils
import torchvision.transforms as transforms
from torch.utils.tensorboard import SummaryWriter
from torch.utils.data import DataLoader
from torchensemble.utils import set_module
from torchensemble.voting import VotingClassifier

classes = ['plane', 'car', 'bird', 'cat', 'deer', 'dog', 'frog', 'horse', 'ship', 'truck']


def get_args_parser(add_help=True):
    import argparse

    parser = argparse.ArgumentParser(description="PyTorch Classification Training", add_help=add_help)

    parser.add_argument("--data-path", default=r"E:\Pytorch-Tutorial-2nd\data\datasets\cifar10-office", type=str,
                        help="dataset path")
    parser.add_argument("--model", default="resnet8", type=str, help="model name")
    parser.add_argument("--device", default="cuda", type=str, help="device (Use cuda or cpu Default: cuda)")
    parser.add_argument(
        "-b", "--batch-size", default=128, type=int, help="images per gpu, the total batch size is $NGPU x batch_size"
    )
    parser.add_argument("--epochs", default=200, type=int, metavar="N", help="number of total epochs to run")
    parser.add_argument(
        "-j", "--workers", default=4, type=int, metavar="N", help="number of data loading workers (default: 16)"
    )
    parser.add_argument("--opt", default="SGD", type=str, help="optimizer")
    parser.add_argument("--random-seed", default=42, type=int, help="random seed")
    parser.add_argument("--lr", default=0.1, type=float, help="initial learning rate")
    parser.add_argument("--momentum", default=0.9, type=float, metavar="M", help="momentum")
    parser.add_argument(
        "--wd",
        "--weight-decay",
        default=1e-4,
        type=float,
        metavar="W",
        help="weight decay (default: 1e-4)",
        dest="weight_decay",
    )
    parser.add_argument("--lr-step-size", default=80, type=int, help="decrease lr every step-size epochs")
    parser.add_argument("--lr-gamma", default=0.1, type=float, help="decrease lr by a factor of lr-gamma")
    parser.add_argument("--print-freq", default=80, type=int, help="print frequency")
    parser.add_argument("--output-dir", default="./Result", type=str, help="path to save outputs")
    parser.add_argument("--resume", default="", type=str, help="path of checkpoint")
    parser.add_argument("--start-epoch", default=0, type=int, metavar="N", help="start epoch")

    return parser


def main():
    args = get_args_parser().parse_args()
    utils.setup_seed(args.random_seed)
    args.device = torch.device("cuda" if torch.cuda.is_available() else "cpu")

    device = args.device
    data_dir = args.data_path
    result_dir = args.output_dir
    # ------------------------------------  log ------------------------------------
    logger, log_dir = utils.make_logger(result_dir)
    writer = SummaryWriter(log_dir=log_dir)

    # ------------------------------------ step1: dataset ------------------------------------

    normMean = [0.4948052, 0.48568845, 0.44682974]
    normStd = [0.24580306, 0.24236229, 0.2603115]
    normTransform = transforms.Normalize(normMean, normStd)
    train_transform = transforms.Compose([
        transforms.Resize(32),
        transforms.RandomCrop(32, padding=4),
        transforms.ToTensor(),
        normTransform
    ])

    valid_transform = transforms.Compose([
        transforms.ToTensor(),
        normTransform
    ])

    # root变量下需要存放cifar-10-python.tar.gz 文件
    # cifar-10-python.tar.gz可从 "https://www.cs.toronto.edu/~kriz/cifar-10-python.tar.gz" 下载
    train_set = torchvision.datasets.CIFAR10(root=data_dir, train=True, transform=train_transform, download=True)
    test_set = torchvision.datasets.CIFAR10(root=data_dir, train=False, transform=valid_transform, download=True)

    # 构建DataLoder
    train_loader = DataLoader(dataset=train_set, batch_size=args.batch_size, shuffle=True, num_workers=args.workers)
    valid_loader = DataLoader(dataset=test_set, batch_size=args.batch_size, num_workers=args.workers)

    # ------------------------------------ tep2: model ------------------------------------
    model_base = utils.resnet20()
    # model_base = utils.LeNet5()
    model = MyEnsemble(estimator=model_base, n_estimators=3, logger=logger, device=device, args=args,
                       classes=classes, writer=writer, save_dir=log_dir)
    model.set_optimizer(args.opt, lr=args.lr, weight_decay=args.weight_decay)
    model.fit(train_loader, test_loader=valid_loader, epochs=args.epochs)


class MyEnsemble(VotingClassifier):
    def __init__(self, **kwargs):
        # logger, device, args, classes, writer
        super(VotingClassifier, self).__init__(kwargs["estimator"], kwargs["n_estimators"])
        self.logger = kwargs["logger"]
        self.writer = kwargs["writer"]
        self.device = kwargs["device"]
        self.args = kwargs["args"]
        self.classes = kwargs["classes"]
        self.save_dir = kwargs["save_dir"]

    @staticmethod
    def save(model, save_dir, logger):
        """Implement model serialization to the specified directory."""
        if save_dir is None:
            save_dir = "./"

        if not os.path.isdir(save_dir):
            os.mkdir(save_dir)

        # Decide the base estimator name
        if isinstance(model.base_estimator_, type):
            base_estimator_name = model.base_estimator_.__name__
        else:
            base_estimator_name = model.base_estimator_.__class__.__name__

        # {Ensemble_Model_Name}_{Base_Estimator_Name}_{n_estimators}
        filename = "{}_{}_{}_ckpt.pth".format(
            type(model).__name__,
            base_estimator_name,
            model.n_estimators,
        )

        # The real number of base estimators in some ensembles is not same as
        # `n_estimators`.
        state = {
            "n_estimators": len(model.estimators_),
            "model": model.state_dict(),
            "_criterion": model._criterion,
        }
        save_dir = os.path.join(save_dir, filename)

        logger.info("Saving the model to `{}`".format(save_dir))

        # Save
        torch.save(state, save_dir)

        return

    def fit(self, train_loader, epochs=100, log_interval=100, test_loader=None, save_model=True, save_dir=None, ):

        # 模型、优化器、学习率调整器、评估器 列表创建
        estimators = []
        for _ in range(self.n_estimators):
            estimators.append(self._make_estimator())

        optimizers = []
        schedulers = []
        for i in range(self.n_estimators):
            optimizers.append(set_module.set_optimizer(estimators[i],
                                                       self.optimizer_name, **self.optimizer_args))
            scheduler_ = torch.optim.lr_scheduler.MultiStepLR(optimizers[i], milestones=[100, 150],
                                                              gamma=self.args.lr_gamma)  # 设置学习率下降策略
            # scheduler_ = torch.optim.lr_scheduler.StepLR(optimizers[i], step_size=self.args.lr_step_size,
            #                                             gamma=self.args.lr_gamma)  # 设置学习率下降策略
            schedulers.append(scheduler_)

        acc_metrics = []
        for i in range(self.n_estimators):
            # task类型与任务一致
            # num_classes与分类任务的类别数一致
            acc_metrics.append(torchmetrics.Accuracy(task="multiclass", num_classes=len(self.classes)))

        self._criterion = nn.CrossEntropyLoss()

        # epoch循环迭代
        best_acc = 0.
        for epoch in range(epochs):

            # training
            for model_idx, (estimator, optimizer, scheduler) in enumerate(zip(estimators, optimizers, schedulers)):
                loss_m_train, acc_m_train, mat_train = \
                    utils.ModelTrainerEnsemble.train_one_epoch(
                        train_loader, estimator, self._criterion, optimizer, scheduler, epoch,
                        self.device, self.args, self.logger, self.classes)
                # 学习率更新
                scheduler.step()

                # 记录
                self.writer.add_scalars('Loss_group', {'train_loss_{}'.format(model_idx):
                                                           loss_m_train.avg}, epoch)
                self.writer.add_scalars('Accuracy_group', {'train_acc_{}'.format(model_idx):
                                                               acc_m_train.avg}, epoch)
                self.writer.add_scalar('learning rate', scheduler.get_last_lr()[0], epoch)
                # 训练混淆矩阵图
                conf_mat_figure_train = utils.show_conf_mat(mat_train, classes, "train", save_dir, epoch=epoch,
                                                            verbose=epoch == epochs - 1, save=False)
                self.writer.add_figure('confusion_matrix_train', conf_mat_figure_train, global_step=epoch)

            # validate
            loss_valid_meter, acc_valid, top1_group, mat_valid = \
                utils.ModelTrainerEnsemble.evaluate(test_loader, estimators, self._criterion, self.device, self.classes)

            # 日志
            self.writer.add_scalars('Loss_group', {'valid_loss':
                                                       loss_valid_meter.avg}, epoch)
            self.writer.add_scalars('Accuracy_group', {'valid_acc':
                                                           acc_valid * 100}, epoch)
            # 验证混淆矩阵图
            conf_mat_figure_valid = utils.show_conf_mat(mat_valid, classes, "valid", save_dir, epoch=epoch,
                                                        verbose=epoch == epochs - 1, save=False)
            self.writer.add_figure('confusion_matrix_valid', conf_mat_figure_valid, global_step=epoch)

            self.logger.info(
                'Epoch: [{:0>3}/{:0>3}]  '
                'Train Loss avg: {loss_train:>6.4f}  '
                'Valid Loss avg: {loss_valid:>6.4f}  '
                'Train Acc@1 avg:  {top1_train:>7.2f}%   '
                'Valid Acc@1 avg: {top1_valid:>7.2%}    '
                'LR: {lr}'.format(
                    epoch, self.args.epochs, loss_train=loss_m_train.avg, loss_valid=loss_valid_meter.avg,
                    top1_train=acc_m_train.avg, top1_valid=acc_valid, lr=schedulers[0].get_last_lr()[0]))

            for model_idx, top1_meter in enumerate(top1_group):
                self.writer.add_scalars('Accuracy_group',
                                        {'valid_acc_{}'.format(model_idx): top1_meter.compute() * 100}, epoch)

            if acc_valid > best_acc:
                best_acc = acc_valid
                self.estimators_ = nn.ModuleList()
                self.estimators_.extend(estimators)
                if save_model:
                    self.save(self, self.save_dir, self.logger)


if __name__ == "__main__":
    main()

效果图

本实验采用3个学习器进行投票式集成，因此绘制了7条曲线，其中各学习器在训练和验证各有2条曲线，集成模型的结果通过 valid_acc输出（蓝色），通过下图可发现，集成模型与三个基学习器相比，分类准确率都能提高3-4百分点左右，是非常高的提升了。

参考

7.7 TorchEnsemble 模型集成库 · PyTorch实用教程（第二版） (tingsongyu.github.io)