基于交通灯数据集的端到端分类

本文介绍如何使用pytorch实现一个基于交通灯数据集的分类系统。

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
torchsummary：pip 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_Classification，git 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：

数据集的可通过如下链接进行下载：onedrive，baiduyun，google。
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：

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__和__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(\'\n\')


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：

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！