目录
1. 前言
2. vgg 的水果分类
2.1 训练
2.2 训练结果
2.3 推理
1. 前言
VGG网络是牛津大学研究人员提出的一种卷积神经网络（CNN）架构。
它被广泛用于图像分类和特征提取任务。VGG网络由一系列卷积层和完全连接的层组成。网络架构的特点是简单，整个网络中只使用了3x3卷积滤波器和最大池化层。VGG模型以其深度架构而闻名，VGG16和VGG19等变体分别具有16层和19层。VGG网络对后续CNN架构的发展产生了影响，并在ImageNet等图像分类基准上取得了优异的性能。
vgg网络模型的亮点：
- 以连续的3*3卷积替代更大的卷积核（7*7），经过证明，这种小的连续卷积核实现的感受野和较大的卷积一样，并且计算量要少得多
- 更简单的模型架构，完全由卷积+下采样实现，避免了以往杂乱无章的网络架构
值得一提的是，vgg模型的效果非常出色，不过因为网络很臃肿导致完全训练不动，这也是vgg最大的鸡肋。
本人认为，vgg模型还是比较成功的，首先是连续的3*3卷积可以替代更大的卷积核。几乎从vgg以后，卷积核全部换成了3*3。并且vgg模型的简单，也为后续的改进提供思路，像这种简单的卷积、下采样的堆叠就可以实现更好的效果，虽然后续的resnet提出了shortcut的瓶颈结构，但原理其实也是下采样一半，卷积核通道数翻倍，就是vgg架构的变迁而已
2. vgg 的水果分类
项目下载：图像识别项目：vgg系列网络(vgg11、vgg13、vgg16等)实现的迁移学习、图像识别项目：33种水果图像分类资源-CSDN文库
其中，data是数据集，inference用于推理图像，runs保存训练生成的结果，train用于训练，predict用于推理，utils是需要的工具函数
这里没有提供单独的验证脚本，因为本人习惯，集成在train里面，在训练的时候一起评估了
2.1 训练
水果数据集经过处理如下：
这里是33类别的水果分类，训练集约有11k张图片，验证集约有5k张图片
标签保存在字典里（代码会自动生成）
{"0": "Apple Braeburn","1": "Apple Granny Smith","2": "Apricot","3": "Avocado","4": "Banana","5": "Blueberry","6": "Cactus fruit","7": "Cantaloupe","8": "Cherry","9": "Clementine","10": "Corn","11": "Cucumber Ripe","12": "Grape Blue","13": "Kiwi","14": "Lemon","15": "Limes","16": "Mango","17": "Onion White","18": "Orange","19": "Papaya","20": "Passion Fruit","21": "Peach","22": "Pear","23": "Pepper Green","24": "Pepper Red","25": "Pineapple","26": "Plum","27": "Pomegranate","28": "Potato Red","29": "Raspberry","30": "Strawberry","31": "Tomato","32": "Watermelon"
}2.2 训练结果
本人用vgg11进行训练，效果太好了，直接100%准确率....
训练可以选择的指标如下：
parser.add_argument("--model", default='vgg11', type=str,help='vgg11、vgg13、vgg16、vgg19、vgg11_bn、vgg13_bn、vgg16_bn、vgg19_bn')parser.add_argument("--pretrained", default=True, type=bool) # 采用官方权重parser.add_argument("--freeze_layers", default=True, type=bool) # 冻结权重parser.add_argument("--batch-size", default=4, type=int)parser.add_argument("--epochs", default=10, type=int)parser.add_argument("--optim", default='SGD', type=str,help='SGD、Adam') # 优化器选择parser.add_argument('--lr', default=0.001, type=float)parser.add_argument('--lrf',default=0.0001,type=float) # 最终学习率 = lr * lrf需要注意的是，网络的输出是经过更改的，代码会根据数据集自动生成num classes，不需要自行设置
tmp = net.classifier[3].out_featuresnet.classifier[6] = torch.nn.Linear(tmp,num,bias=True)结果在runs里面下：
这里的训练日志全部在log的json文件中，参考如下：
{"train parameters": {"model": "vgg11","pretrained": true,"freeze_layers": true,"batch_size": 4,"epochs": 10,"optim": "SGD","lr": 0.001,"lrf": 0.0001},"total paramerters": 128901537,"train paramerters": 119681057,"epoch:0": {"train info": {"accuracy": 0.9742591024547211,"Apple Braeburn": {"Precision": 0.948,"Recall": 0.9507,"Specificity": 0.9984,"F1 score": 0.9493},"Apple Granny Smith": {"Precision": 0.9883,"Recall": 0.9768,"Specificity": 0.9997,"F1 score": 0.9825},"Apricot": {"Precision": 0.9791,"Recall": 0.9507,"Specificity": 0.9994,"F1 score": 0.9647},"Avocado": {"Precision": 0.9691,"Recall": 0.9431,"Specificity": 0.9992,"F1 score": 0.9559},"Banana": {"Precision": 0.9912,"Recall": 0.9883,"Specificity": 0.9997,"F1 score": 0.9897},"Blueberry": {"Precision": 0.9753,"Recall": 0.9753,"Specificity": 0.9993,"F1 score": 0.9753},"Cactus fruit": {"Precision": 0.9797,"Recall": 0.9825,"Specificity": 0.9994,"F1 score": 0.9811},"Cantaloupe": {"Precision": 0.9885,"Recall": 0.9942,"Specificity": 0.9997,"F1 score": 0.9913},"Cherry": {"Precision": 0.9883,"Recall": 0.9768,"Specificity": 0.9997,"F1 score": 0.9825},"Clementine": {"Precision": 0.9218,"Recall": 0.9621,"Specificity": 0.9976,"F1 score": 0.9415},"Corn": {"Precision": 1.0,"Recall": 0.9905,"Specificity": 1.0,"F1 score": 0.9952},"Cucumber Ripe": {"Precision": 0.9926,"Recall": 0.9782,"Specificity": 0.9998,"F1 score": 0.9853},"Grape Blue": {"Precision": 0.9717,"Recall": 0.9956,"Specificity": 0.9982,"F1 score": 0.9835},"Kiwi": {"Precision": 0.9698,"Recall": 0.9817,"Specificity": 0.9991,"F1 score": 0.9757},"Lemon": {"Precision": 0.971,"Recall": 0.971,"Specificity": 0.9991,"F1 score": 0.971},"Limes": {"Precision": 0.9443,"Recall": 0.9883,"Specificity": 0.9983,"F1 score": 0.9658},"Mango": {"Precision": 0.9561,"Recall": 0.9534,"Specificity": 0.9987,"F1 score": 0.9547},"Onion White": {"Precision": 0.9741,"Recall": 0.9805,"Specificity": 0.9993,"F1 score": 0.9773},"Orange": {"Precision": 0.9852,"Recall": 0.9911,"Specificity": 0.9996,"F1 score": 0.9881},"Papaya": {"Precision": 0.9431,"Recall": 0.913,"Specificity": 0.9983,"F1 score": 0.9278},"Passion Fruit": {"Precision": 0.9941,"Recall": 0.9767,"Specificity": 0.9998,"F1 score": 0.9853},"Peach": {"Precision": 0.9263,"Recall": 0.9478,"Specificity": 0.9977,"F1 score": 0.9369},"Pear": {"Precision": 0.9754,"Recall": 0.9754,"Specificity": 0.9989,"F1 score": 0.9754},"Pepper Green": {"Precision": 0.9936,"Recall": 0.9936,"Specificity": 0.9998,"F1 score": 0.9936},"Pepper Red": {"Precision": 0.9789,"Recall": 0.9936,"Specificity": 0.9991,"F1 score": 0.9862},"Pineapple": {"Precision": 0.9971,"Recall": 0.9913,"Specificity": 0.9999,"F1 score": 0.9942},"Plum": {"Precision": 0.9904,"Recall": 0.984,"Specificity": 0.9997,"F1 score": 0.9872},"Pomegranate": {"Precision": 0.9531,"Recall": 0.942,"Specificity": 0.9986,"F1 score": 0.9475},"Potato Red": {"Precision": 0.9773,"Recall": 0.9556,"Specificity": 0.9994,"F1 score": 0.9663},"Raspberry": {"Precision": 1.0,"Recall": 0.9883,"Specificity": 1.0,"F1 score": 0.9941},"Strawberry": {"Precision": 0.9741,"Recall": 0.9826,"Specificity": 0.9992,"F1 score": 0.9783},"Tomato": {"Precision": 0.9728,"Recall": 0.9691,"Specificity": 0.9988,"F1 score": 0.9709},"Watermelon": {"Precision": 0.997,"Recall": 0.982,"Specificity": 0.9999,"F1 score": 0.9894},"mean precision": 0.9747666666666668,"mean recall": 0.9735090909090911,"mean specificity": 0.9991909090909091,"mean f1 score": 0.9740454545454547},"valid info": {"accuracy": 0.9924662965880403,"Apple Braeburn": {"Precision": 0.9671,"Recall": 1.0,"Specificity": 0.999,"F1 score": 0.9833},"Apple Granny Smith": {"Precision": 1.0,"Recall": 1.0,"Specificity": 1.0,"F1 score": 1.0},"Apricot": {"Precision": 0.9545,"Recall": 1.0,"Specificity": 0.9986,"F1 score": 0.9767},"Avocado": {"Precision": 1.0,"Recall": 1.0,"Specificity": 1.0,"F1 score": 1.0},"Banana": {"Precision": 1.0,"Recall": 1.0,"Specificity": 1.0,"F1 score": 1.0},"Blueberry": {"Precision": 1.0,"Recall": 1.0,"Specificity": 1.0,"F1 score": 1.0},"Cactus fruit": {"Precision": 1.0,"Recall": 1.0,"Specificity": 1.0,"F1 score": 1.0},"Cantaloupe": {"Precision": 1.0,"Recall": 1.0,"Specificity": 1.0,"F1 score": 1.0},"Cherry": {"Precision": 1.0,"Recall": 1.0,"Specificity": 1.0,"F1 score": 1.0},"Clementine": {"Precision": 1.0,"Recall": 1.0,"Specificity": 1.0,"F1 score": 1.0},"Corn": {"Precision": 1.0,"Recall": 1.0,"Specificity": 1.0,"F1 score": 1.0},"Cucumber Ripe": {"Precision": 1.0,"Recall": 1.0,"Specificity": 1.0,"F1 score": 1.0},"Grape Blue": {"Precision": 1.0,"Recall": 1.0,"Specificity": 1.0,"F1 score": 1.0},"Kiwi": {"Precision": 0.9789,"Recall": 1.0,"Specificity": 0.9994,"F1 score": 0.9893},"Lemon": {"Precision": 0.9735,"Recall": 1.0,"Specificity": 0.9992,"F1 score": 0.9866},"Limes": {"Precision": 0.9932,"Recall": 1.0,"Specificity": 0.9998,"F1 score": 0.9966},"Mango": {"Precision": 1.0,"Recall": 0.9932,"Specificity": 1.0,"F1 score": 0.9966},"Onion White": {"Precision": 1.0,"Recall": 0.9924,"Specificity": 1.0,"F1 score": 0.9962},"Orange": {"Precision": 1.0,"Recall": 1.0,"Specificity": 1.0,"F1 score": 1.0},"Papaya": {"Precision": 0.9859,"Recall": 0.9524,"Specificity": 0.9996,"F1 score": 0.9689},"Passion Fruit": {"Precision": 1.0,"Recall": 1.0,"Specificity": 1.0,"F1 score": 1.0},"Peach": {"Precision": 0.9074,"Recall": 1.0,"Specificity": 0.9969,"F1 score": 0.9515},"Pear": {"Precision": 0.9951,"Recall": 0.9808,"Specificity": 0.9998,"F1 score": 0.9879},"Pepper Green": {"Precision": 1.0,"Recall": 1.0,"Specificity": 1.0,"F1 score": 1.0},"Pepper Red": {"Precision": 1.0,"Recall": 1.0,"Specificity": 1.0,"F1 score": 1.0},"Pineapple": {"Precision": 1.0,"Recall": 1.0,"Specificity": 1.0,"F1 score": 1.0},"Plum": {"Precision": 1.0,"Recall": 1.0,"Specificity": 1.0,"F1 score": 1.0},"Pomegranate": {"Precision": 1.0,"Recall": 0.9252,"Specificity": 1.0,"F1 score": 0.9611},"Potato Red": {"Precision": 1.0,"Recall": 0.8963,"Specificity": 1.0,"F1 score": 0.9453},"Raspberry": {"Precision": 1.0,"Recall": 1.0,"Specificity": 1.0,"F1 score": 1.0},"Strawberry": {"Precision": 1.0,"Recall": 1.0,"Specificity": 1.0,"F1 score": 1.0},"Tomato": {"Precision": 1.0,"Recall": 1.0,"Specificity": 1.0,"F1 score": 1.0},"Watermelon": {"Precision": 1.0,"Recall": 1.0,"Specificity": 1.0,"F1 score": 1.0},"mean precision": 0.9925939393939395,"mean recall": 0.9921303030303031,"mean specificity": 0.9997666666666667,"mean f1 score": 0.992121212121212}},
仅仅跑了几轮，已经100准确率了
loss和acc曲线：
其他的评估指标曲线：
训练集和测试集的混淆矩阵：
2.3 推理
推理需要predict脚本，设定的参数在下面，model要保证和训练的版本一样
parser.add_argument("--model", default='vgg11', type=str,help='vgg11、vgg13、vgg16、vgg19、vgg11_bn、vgg13_bn、vgg16_bn、vgg19_bn')parser.add_argument("--weights", default='runs/weights/best.pth', type=str, help='best、last')
只需要把想要推理的数据放在infer_img下即可
运行即可推理 ：
