zoukankan      html  css  js  c++  java
  • 基于交通灯数据集的端到端分类

    抓住11月的尾巴,这里写上昨天做的一个DL的作业吧,作业很简单,基于交通灯的图像分类,但这确是让你从0构建深度学习系统的好例子,很多已有的数据集都封装好了,直接调用,这篇文章将以pytorch这个深度学习框架一步步搭建分类系统。

    软件包要求:
    pytorch:0.4.0
    torchsummarypip install torchsummary
    cv2: pip install opencv-python
    matplotlib
    numpy

    所有代码托管到github上,链接如下:https://github.com/FangYang970206/TL_Dataset_Classificationgit clone https://github.com/FangYang970206/TL_Dataset_Classification到本地。

    1.数据集简介

    数据集有10个类别,分别是红灯的圆球,向左,向右,向上和负例以及绿灯的圆球,向左,向右,向上和负例,如下图所示:
    image_1cthr3cpb93e1ia7dnl1kl51tjh9.png-227.7kB
    数据集的可通过如下链接进行下载:onedrivebaiduyungoogle
    下完数据集后,解压到文件夹TL_Dataset_Classification-master中,得到一个新的文件夹TL_Dataset,可以看到TL_Dataset有以下目录:
    image_1cthuhco8l4f1n05hen27a1ou216.png-14.6kB

    2.代码实战

    代码是在vscode上编写的,支持flask8,总共有9个文件,下面一一介绍。建议在看代码的时候从main.py文件开始看,大致脉络就清楚了。

    2.1 model.py

    对于一个深度学习系统来说,model应该是最初的想法,我们想构造什么样的模型来拟合数据集,所以先写model,代码如下:

    import torch.nn as nn
    from torchsummary import summary
    
    
    class A2NN(nn.Module):
        def __init__(self, ):
            super(A2NN, self).__init__()
            self.main = nn.Sequential(
                nn.Conv2d(3, 16, 3, 1, 1),
                nn.BatchNorm2d(16),
                nn.ReLU(inplace=True),
                nn.Conv2d(16, 32, 3, 1, 1),
                nn.MaxPool2d(2, 2),
                nn.BatchNorm2d(32),
                nn.ReLU(inplace=True),
                nn.Conv2d(32, 32, 3, 1, 1),
                nn.MaxPool2d(2, 2),
                nn.BatchNorm2d(32),
                nn.ReLU(inplace=True),
                nn.Conv2d(32, 64, 3, 1, 1),
                nn.MaxPool2d(2, 2),
                nn.BatchNorm2d(64),
                nn.ReLU(inplace=True),
            )
            self.linear = nn.Linear(4*4*64, 9)
    
        def forward(self, inp):
            x = self.main(inp)
            x = x.view(x.shape[0], -1)
            x = self.linear(x)
            return x
    
    
    if __name__ == "__main__":
        nn = A2NN()
        summary(nn, (3, 32, 32))
    

    model代码不复杂,很简单,这里不多介绍,缺少基础的朋友还请自行补基础。

    2.2 dataset.py

    第二步我们要构建数据集类,pytorch封装了一个torch.utils.data.Dataset的类,我们可以重载__len____getitem__方法,来得到自己的数据集管道,__len__方法是返回数据集的长度,__getitem__是支持从0到len(self)互斥范围内的整数索引,返回的是索引对应的数据和标签。代码如下:

    import torch
    import cv2
    import torch.utils.data as data
    
    
    class_light = {
                    'Red Circle': 0,
                    'Green Circle': 1,
                    'Red Left': 2,
                    'Green Left': 3,
                    'Red Up': 4,
                    'Green Up': 5,
                    'Red Right': 6,
                    'Green Right': 7,
                    'Red Negative': 8,
                    'Green Negative': 8
    }
    
    
    class Traffic_Light(data.Dataset):
        def __init__(self, dataset_names, img_resize_shape):
            super(Traffic_Light, self).__init__()
            self.dataset_names = dataset_names
            self.img_resize_shape = img_resize_shape
    
        def __getitem__(self, ind):
            img = cv2.imread(self.dataset_names[ind])
            img = cv2.resize(img, self.img_resize_shape)
            img = img.transpose(2, 0, 1)-127.5/127.5
            for key in class_light.keys():
                if key in self.dataset_names[ind]:
                    label = class_light[key]
            # pylint: disable=E1101,E1102
            return torch.from_numpy(img), torch.tensor(label)
            # pylint: disable=E1101,E1102
    
        def __len__(self):
            return len(self.dataset_names)
    
    
    if __name__ == '__main__':
        from torch.utils.data import DataLoader
        from glob import glob
        import os
    
        path = 'TL_Dataset/Green Up/'
        names = glob(os.path.join(path, '*.png'))
        dataset = Traffic_Light(names, (32, 32))
        dataload = DataLoader(dataset, batch_size=1)
        for ind, (inp, label) in enumerate(dataload):
            print("{}-inp_size:{}-label_size:{}".format(ind, inp.numpy().shape,
                                                        label.numpy().shape))
    

    2.3 util.py

    在上面的dataset.py中,class初始化时,传入了dataset_names,所以utils.py文件中就通过get_train_val_names函数得到训练数据集和验证数据集的names,还有一个函数是检查文件夹是否存在,不存在建立文件夹。代码如下:

    import os
    from glob import glob
    
    
    def get_train_val_names(dataset_path, remove_names, radio=0.3):
        train_names = []
        val_names = []
        dataset_paths = os.listdir(dataset_path)
        for n in remove_names:
            dataset_paths.remove(n)
        for path in dataset_paths:
            sub_dataset_path = os.path.join(dataset_path, path)
            sub_dataset_names = glob(os.path.join(sub_dataset_path, '*.png'))
            sub_dataset_len = len(sub_dataset_names)
            val_names.extend(sub_dataset_names[:int(radio*sub_dataset_len)])
            train_names.extend(sub_dataset_names[int(radio*sub_dataset_len):])
        return {'train': train_names, 'val': val_names}
    
    
    def check_folder(path):
        if not os.path.exists(path):
            os.mkdir(path)
    

    2.4 trainer.py

    model构造好了,数据集也准备好了,现在就需要准备如果训练了,这就是trainer.py文件的作用,trainer.py构建了Trainer类,通过传入训练的一系列参数,调用Trainer.train函数进行训练,并返回loss,代码如下:

    import torch.nn as nn
    from torch.optim import Adam
    
    
    class Trainer:
        def __init__(self, model, dataload, epoch, lr, device):
            self.model = model
            self.dataload = dataload
            self.epoch = epoch
            self.lr = lr
            self.device = device
            self.optimizer = Adam(self.model.parameters(), lr=self.lr)
            self.criterion = nn.CrossEntropyLoss().to(self.device)
    
        def __epoch(self, epoch):
            self.model.train()
            loss_sum = 0
            for ind, (inp, label) in enumerate(self.dataload):
                inp = inp.float().to(self.device)
                label = label.long().to(self.device)
                self.optimizer.zero_grad()
                out = self.model.forward(inp)
                loss = self.criterion(out, label)
                loss.backward()
                loss_sum += loss.item()
                self.optimizer.step()
                print('epoch{}_step{}_train_loss_: {}'.format(epoch,
                                                              ind,
                                                              loss.item()))
            return loss_sum/(ind+1)
    
        def train(self):
            train_loss = self.__epoch(self.epoch)
            return train_loss
    

    2.5 validator.py

    trainer.py文件是用来进行训练数据集的,训练过程中,我们是需要有验证集来判断我们模型的训练效果,所以这里有validator.py文件,里面封装了Validator类,与Trainer.py类似,但不同的是,我们不训练,不更新参数,model处于eval模式,代码上会有一些跟Trainer不一样,通过调用Validator.eval函数返回loss,代码如下:

    import torch.nn as nn
    
    
    class Validator:
        def __init__(self, model, dataload, epoch, device, batch_size):
            self.model = model
            self.dataload = dataload
            self.epoch = epoch
            self.device = device
            self.batch_size = batch_size
            self.criterion = nn.CrossEntropyLoss().to(self.device)
    
        def __epoch(self, epoch):
            self.model.eval()
            loss_sum = 0
            for ind, (inp, label) in enumerate(self.dataload):
                inp = inp.float().to(self.device)
                label = label.long().to(self.device)
                out = self.model.forward(inp)
                loss = self.criterion(out, label)
                loss_sum += loss.item()
            return {'val_loss': loss_sum/(ind+1)}
    
        def eval(self):
            val_loss = self.__epoch(self.epoch)
            return val_loss
    

    2.6 logger.py

    我们想看整个学习的过程,可以通过看学习曲线来进行观察。所以这里写了一个logger.py文件,用来对训练loss和验证loss进行统计并画图。代码如下:

    import matplotlib.pyplot as plt
    import os
    
    
    class Logger:
        def __init__(self, save_path):
            self.save_path = save_path
    
        def update(self, Kwarg):
            self.__plot(Kwarg)
    
        def __plot(self, Kwarg):
            save_img_path = os.path.join(self.save_path, 'learning_curve.png')
            plt.clf()
            plt.plot(Kwarg['train_losses'], label='Train', color='g')
            plt.plot(Kwarg['val_losses'], label='Val', color='b')
            plt.xlabel('epoch')
            plt.ylabel('loss')
            plt.legend()
            plt.title('learning_curve')
            plt.savefig(save_img_path)
    

    2.7 main.py

    main.py文件将上面所有的东西结合到一起,代码如下:

    import torch
    import argparse
    
    from model import A2NN
    from dataset import Traffic_Light
    from utils import get_train_val_names, check_folder
    from trainer import Trainer
    from validator import Validator
    from logger import Logger
    from torch.utils.data import DataLoader
    
    
    def main():
        parse = argparse.ArgumentParser()
        parse.add_argument('--dataset_path', type=str, default='TL_Dataset/')
        parse.add_argument('--remove_names', type=list, default=['README.txt',
                                                                 'README.png',
                                                                 'Testset'])
        parse.add_argument('--img_resize_shape', type=tuple, default=(32, 32))
        parse.add_argument('--batch_size', type=int, default=1024)
        parse.add_argument('--lr', type=float, default=0.001)
        parse.add_argument('--num_workers', type=int, default=4)
        parse.add_argument('--epochs', type=int, default=200)
        parse.add_argument('--val_size', type=float, default=0.3)
        parse.add_argument('--save_model', type=bool, default=True)
        parse.add_argument('--save_path', type=str, default='logs/')
    
        args = vars(parse.parse_args())
    
        check_folder(args['save_path'])
    
        # pylint: disable=E1101
        device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
        # pylint: disable=E1101
    
        model = A2NN().to(device)
    
        names = get_train_val_names(args['dataset_path'], args['remove_names'])
    
        train_dataset = Traffic_Light(names['train'], args['img_resize_shape'])
        val_dataset = Traffic_Light(names['val'], args['img_resize_shape'])
    
        train_dataload = DataLoader(train_dataset,
                                    batch_size=args['batch_size'],
                                    shuffle=True,
                                    num_workers=args['num_workers'])
    
        val_dataload = DataLoader(val_dataset,
                                  batch_size=args['batch_size'],
                                  shuffle=True,
                                  num_workers=args['num_workers'])
    
        loss_logger = Logger(args['save_path'])
    
        logger_dict = {'train_losses': [],
                       'val_losses': []}
    
        for epoch in range(args['epochs']):
            print('<Main> epoch{}'.format(epoch))
            trainer = Trainer(model, train_dataload, epoch, args['lr'], device)
            train_loss = trainer.train()
            if args['save_model']:
                state = model.state_dict()
                torch.save(state, 'logs/nn_state.t7')
            validator = Validator(model, val_dataload, epoch,
                                  device, args['batch_size'])
            val_loss = validator.eval()
            logger_dict['train_losses'].append(train_loss)
            logger_dict['val_losses'].append(val_loss['val_loss'])
    
            loss_logger.update(logger_dict)
    
    
    if __name__ == '__main__':
        main()
    

    2.8 compute_prec.py和submit.py

    其实上面的七个文件,已经是结束了,下面两个文件一个是用来计算精确度的,一个是用来提交答案的。有兴趣可以看看。
    compute_prec.py代码如下:

    import torch
    import numpy as np
    import argparse
    
    from model import A2NN
    from dataset import Traffic_Light
    from torch.utils.data import DataLoader
    from utils import get_train_val_names, check_folder
    
    
    def main():
        parse = argparse.ArgumentParser()
        parse.add_argument('--dataset_path', type=str, default='TL_Dataset/')
        parse.add_argument('--remove_names', type=list, default=['README.txt',
                                                                 'README.png',
                                                                 'Testset'])
        parse.add_argument('--img_resize_shape', type=tuple, default=(32, 32))
        parse.add_argument('--num_workers', type=int, default=4)
        parse.add_argument('--val_size', type=float, default=0.3)
        parse.add_argument('--save_path', type=str, default='logs/')
    
        args = vars(parse.parse_args())
    
        check_folder(args['save_path'])
    
        # pylint: disable=E1101
        device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
        # pylint: disable=E1101
    
        model = A2NN().to(device)
        model.load_state_dict(torch.load(args['save_path']+'nn_state.t7'))
    
        model.eval()
    
        names = get_train_val_names(args['dataset_path'], args['remove_names'])
    
        val_dataset = Traffic_Light(names['val'], args['img_resize_shape'])
    
        val_dataload = DataLoader(val_dataset,
                                  batch_size=1,
                                  num_workers=args['num_workers'])
    
        count = 0
        for ind, (inp, label) in enumerate(val_dataload):
            inp = inp.float().to(device)
            label = label.long().to(device)
            output = model.forward(inp)
            output = np.argmax(output.to('cpu').detach().numpy(), axis=1)
            label = label.to('cpu').numpy()
            count += 1 if output == label else 0
    
        print('precision: {}'.format(count/(ind+1)))
    
    
    if __name__ == "__main__":
        main()
    

    submit.py代码如下:

    import torch
    import numpy as np
    import argparse
    import os
    import cv2
    
    from model import A2NN
    from utils import check_folder
    
    
    def main():
        parse = argparse.ArgumentParser()
        parse.add_argument('--dataset_path', type=str,
                           default='TL_Dataset/Testset/')
        parse.add_argument('--img_resize_shape', type=tuple, default=(32, 32))
        parse.add_argument('--num_workers', type=int, default=4)
        parse.add_argument('--save_path', type=str, default='logs/')
    
        args = vars(parse.parse_args())
    
        check_folder(args['save_path'])
    
        # pylint: disable=E1101
        device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
        # pylint: disable=E1101
    
        model = A2NN().to(device)
        model.load_state_dict(torch.load(args['save_path']+'nn_state.t7'))
    
        model.eval()
    
        txt_path = os.path.join(args['save_path'], 'result.txt')
        with open(txt_path, 'w') as f:
            for i in range(20000):
                name = os.path.join(args['dataset_path'], '{}.png'.format(i))
                img = cv2.imread(name)
                img = cv2.resize(img, args['img_resize_shape'])
                img = img.transpose(2, 0, 1)-127.5/127.5
                img = torch.unsqueeze(torch.from_numpy(img).float(), dim=0)
                img = img.to(device)
                output = model.forward(img).to('cpu').detach().numpy()
                img_class = np.argmax(output, axis=1)
                f.write(name.split('/')[2] + ' ' + str(img_class[0]))
                f.write('
    ')
    
    
    if __name__ == "__main__":
        main()
    

    3. 代码如下运行

    将数据集下载在文件夹TL_Dataset_Classification,解压后,在TL_Dataset_Classification文件中进入终端,运行命令:

    $ python main.py
    

    如果还想计算精确度,在训练玩数据集之后,运行命令:

    $ python compute_prec.py
    

    有运行可以到github上提issue或者在给我的邮箱867540289@qq.com发邮件。

    4. 结果

    学习曲线:
    learning_curve.png-22.8kB
    在测试集中,实现97.425%的精确度。

    5. 总结

    好了,11月的尾巴到此结束,希望能对你学习深度学习问题和pytorch有所帮助。12月马上到,祝我数学考试顺利,也祝各位开开心心!

  • 相关阅读:
    Condtion type Z123 is mandatory!
    PUSU 拆分后发货和开票的时间节点问题
    mass create DN
    Mike Piehl
    变式配置简介 VARIANT CONFIGURATION
    4th Dec 2018
    (转)SQL Server 2012笔记分享-25:配置备份维护计划
    从ICassFactory为CLSID为{17BCA6E8-A950-497E-B2F9-AF6AA475916F}的COM组件创建实例失败,原因是出现以下错误:c001f011.(Microsoft.Server.manageDTS)
    SQL Server 2008 Tempdb 数据库迁移
    .woff HTTP GET 404 (Not Found)
  • 原文地址:https://www.cnblogs.com/fydeblog/p/10046456.html
Copyright © 2011-2022 走看看