基于MobileNetV2的柠檬外观分类实践

本实践基于广岛Quest2020柠檬外观分类赛题，用飞桨2.0搭建卷积神经网络。先解压数据集，用train.csv训练，划分80%为训练集、20%为验证集。经数据预处理、构建数据集、配置visualdl后，选用MobileNetV2模型，以SGD优化器等训练，最终验证集准确率达98%-100%，可通过调优提升性能。

项目的赛题源自比赛广岛Quest2020

柠檬外观分类使用的图像数据(第1阶段)用广岛县的柠檬形象挑战外观分类。

免费影视、动漫、音乐、游戏、小说资源长期稳定更新！ 👉 点此立即查看 👈

比赛链接https://signate.jp/competitions/431

1.任务描述

如何根据据图像的视觉内容为图像赋予一个语义类别是图像分类的目标，也是图像检索、图像内容分析和目标识别等问题的基础。

本实践旨在通过一个柠檬分类的案列，让大家理解和掌握如何使用飞桨2.0搭建一个卷积神经网络。

特别提示：本实践所用数据集均来自互联网，请勿用于商务用途。

解压文件，使用train.csv训练，测试使用分出来的验证集。最后以在验证集上的准确率作为最终分数。

2.调优

思考并动手进行调优，以在验证集上的准确率为评价指标，验证集上准确率越高，得分越高！模型大家可以更换，调参技巧任选，代码需要大家自己调通。

解压数据集

!cd data $$/!unzip -oq /home/aistudio/data/data73045/lemon_homework.zip !unzip -oq /home/aistudio/lemon_homework/lemon_lesson.zip -d /home/aistudio/lemon_homework/!unzip -oq /home/aistudio/lemon_homework/lemon_lesson/test_images.zip -d /home/aistudio/lemon_homework/!unzip -oq /home/aistudio/lemon_homework/lemon_lesson/train_images.zip -d /home/aistudio/lemon_homework/

第一步 观察数据格式 了解赛题目的

题目分析

1.题意 给出了训练集柠檬的图片，让我们输出测试集的图片的规格

2.数据集模样

3.关键信息 第一列为id 图片的路径

第二列为label 图片的等级

以下为题目当中给出的：

有四个等级：0:優良、1:良、2:加工品、3:規格外。

第二步导入必要的库，数据解析并存储

代码逻辑：

导入库->读取数据->打乱数据->划分数据->数据预处理

# 导入所需要的库from sklearn.utils import shuffleimport osimport pandas as pdimport numpy as npfrom PIL import Imageimport paddleimport paddle.nn as nnfrom paddle.io import Datasetimport paddle.vision.transforms as Timport paddle.nn.functional as Ffrom paddle.metric import Accuracy #在python中运行代码经常会遇到的情况是——代码可以正常运行但是会提示警告，有时特别讨厌。#那么如何来控制警告输出呢？其实很简单，python通过调用warnings模块中定义的warn()函数来发出警告。我们可以通过警告过滤器进行控制是否发出警告消息。import warningswarnings.filterwarnings("ignore")# 读取数据train_images = pd.read_csv('lemon_homework/train.csv', usecols=['id','class_num'])# labelshuffling 自定义标签打乱def labelShuffling(dataFrame, groupByName = 'class_num'):    groupDataFrame = dataFrame.groupby(by=[groupByName])    labels = groupDataFrame.size()    print("length of label is ", len(labels))    maxNum = max(labels)    lst = pd.DataFrame()    for i in range(len(labels)):        print("Processing label  :", i)        tmpGroupBy = groupDataFrame.get_group(i)        createdShuffleLabels = np.random.permutation(np.array(range(maxNum))) % labels[i]        print("Num of the label is : ", labels[i])        lst=lst.append(tmpGroupBy.iloc[createdShuffleLabels], ignore_index=True)        print("Done" )    # lst.to_csv('test1.csv', index=False)    return lst# 划分训练集和校验集all_size = len(train_images)# print(all_size)train_size = int(all_size * 0.8)train_image_list = train_images[:train_size]val_image_list = train_images[train_size:]df = labelShuffling(train_image_list)df = shuffle(df)train_image_path_list = df['id'].valueslabel_list = df['class_num'].valueslabel_list = paddle.to_tensor(label_list, dtype='int64')train_label_list = paddle.nn.functional.one_hot(label_list, num_classes=4)val_image_path_list = val_image_list['id'].valuesval_label_list = val_image_list['class_num'].valuesval_label_list = paddle.to_tensor(val_label_list, dtype='int64')val_label_list = paddle.nn.functional.one_hot(val_label_list, num_classes=4)# 定义数据预处理data_transforms = T.Compose([    T.Resize(size=(224, 224)),    T.RandomHorizontalFlip(224),    T.RandomVerticalFlip(224),    T.RandomRotation(90),    T.Transpose(),    # HWC -> CHW    T.Normalize(        mean=[0.31169346, 0.25506335, 0.12432463],      #归一化          std=[0.34042713, 0.29819837, 0.1375536],        to_rgb=True)            #计算过程：output[channel] = (input[channel] - mean[channel]) / std[channel]])

length of label is  4Processing label  : 0Num of the label is :  234DoneProcessing label  : 1Num of the label is :  156DoneProcessing label  : 2Num of the label is :  136DoneProcessing label  : 3Num of the label is :  123Done

第三步 构建Dataset

# 构建Datasetclass MyDataset(paddle.io.Dataset):    """    步骤一：继承paddle.io.Dataset类    """    def __init__(self, train_img_list, val_img_list,train_label_list,val_label_list, mode='train'):        """        步骤二：实现构造函数，定义数据读取方式，划分训练和测试数据集        """        super(MyDataset, self).__init__()        self.img = []        self.label = []        # 借助pandas读csv的库        self.train_images = train_img_list        self.test_images = val_img_list        self.train_label = train_label_list        self.test_label = val_label_list        if mode == 'train':            # 读train_images的数据            for img,la in zip(self.train_images, self.train_label):                self.img.append('lemon_homework/train_images/'+img)                self.label.append(la)        else:            # 读test_images的数据            for img,la in zip(self.test_images, self.test_label):                self.img.append('lemon_homework/train_images/'+img)                self.label.append(la)    def load_img(self, image_path):        # 实际使用时使用Pillow相关库进行图片读取即可，这里我们对数据先做个模拟        image = Image.open(image_path).convert('RGB')        return image    def __getitem__(self, index):        """        步骤三：实现__getitem__方法，定义指定index时如何获取数据，并返回单条数据（训练数据，对应的标签）        """        image = self.load_img(self.img[index])        label = self.label[index]        # label = paddle.to_tensor(label)                return data_transforms(image).astype("float32"), paddle.nn.functional.label_smooth(label)    def __len__(self):        """        步骤四：实现__len__方法，返回数据集总数目        """        return len(self.img)#train_loadertrain_dataset = MyDataset(train_img_list=train_image_path_list, val_img_list=val_image_path_list, train_label_list=train_label_list, val_label_list=val_label_list, mode='train')train_loader = paddle.io.DataLoader(train_dataset, places=paddle.CPUPlace(), batch_size=32, shuffle=True, num_workers=0)#val_loaderval_dataset = MyDataset(train_img_list=train_image_path_list, val_img_list=val_image_path_list, train_label_list=train_label_list, val_label_list=val_label_list, mode='test')val_loader = paddle.io.DataLoader(train_dataset, places=paddle.CPUPlace(), batch_size=32, shuffle=True, num_workers=0)

第四步 配置visualdl

配置完之后，可以在左侧可视化页面添加日志和模型文件。

!rm vdl/vdlrecords.model.logfrom visualdl import LogReader, LogWriterargs={    'logdir':'./vdl',    'file_name':'vdlrecords.model.log',    'iters':0,}write = LogWriter(logdir=args['logdir'], file_name=args['file_name'])#iters 初始化为0iters = args['iters'] #自定义Callbackclass Callbk(paddle.callbacks.Callback):    def __init__(self, write, iters=0):        self.write = write        self.iters = iters    def on_train_batch_end(self, step, logs):        self.iters += 1        #记录loss        self.write.add_scalar(tag="loss",step=self.iters,value=logs['loss'][0])        #记录 accuracy        self.write.add_scalar(tag="acc",step=self.iters,value=logs['acc'])

rm: cannot remove 'vdl/vdlrecords.model.log': No such file or directory

第五步 选择合适的网络并进行模型训练

定义输入->模型封装->定义优化器->配置模型->模型训练与评估

from work.mobilenet import MobileNetV2#定义输入input_define = paddle.static.InputSpec(shape=[-1,3,224,224], dtype="float32", name="img")label_define = paddle.static.InputSpec(shape=[-1,1], dtype="int64", name="label")# 模型封装model_res = MobileNetV2(class_dim=4)model = paddle.Model(model_res,inputs=input_define,labels=label_define)# 定义优化器scheduler = paddle.optimizer.lr.LinearWarmup(        learning_rate=0.1, warmup_steps=20, start_lr=0, end_lr=0.1, verbose=True)optim = paddle.optimizer.SGD(learning_rate=scheduler, parameters=model.parameters())# optim = paddle.optimizer.Adam(learning_rate=0.001, parameters=model.parameters())# 配置模型model.prepare(    optim,    paddle.nn.CrossEntropyLoss(soft_label=True),    Accuracy()    )# 模型训练与评估model.fit(        train_loader,        val_loader,        epochs=50,        callbacks=Callbk(write=write, iters=iters),        verbose=1,        batch_size=64,        save_dir="/home/aistudio/iterhui/" #把模型参数、优化器参数保存至自定义的文件夹        )

模型验证

#模型保存model.save('Hapi_MyCNN', False)  # save for inferenceresult = model.evaluate(val_loader,batch_size=64,log_freq=100, verbose=1, num_workers=0, callbacks=Callbk(write=write, iters=iters))print(result)

Eval begin...The loss value printed in the log is the current batch, and the metric is the average value of previous step.step 30/30 [==============================] - loss: 0.4024 - acc: 0.9979 - 642ms/stepEval samples: 936{'loss': [0.40243444], 'acc': 0.9978632478632479}

验证集准确率多次运行大概在98%-100%。