中国一级保护动物识别

!unzip -oq /home/aistudio/data/data63087/动物1-20.zip -d work/

import os

dirpath = "work/动物1-20"# 先得到总的txt后续再进行划分，因为要划分出验证集，所以要先打乱，因为原本是有序的def get_all_txt():
    all_list = []    # 存储数据---图片路径
    data = []    # 存储标签---数字0-20
    label = []    # 记录编号对应的物种
    label_list = []
    i = 0
    j = -1
    for root,dirs,files in os.walk(dirpath): # 分别代表根目录、文件夹、文件
        category_name = ''
        for file in files:
            i = i + 1
            # 文件中每行格式： 图像相对路径      图像的label_id（注意：中间有空格）。              
            # work/动物1-20/蜂猴/b3c393be266ba8182ee2abead6c4bdf6.jpg 0
            imgpath = os.path.join(root,file)
            category_name = imgpath.split("/")[-2]
            all_list.append(imgpath+" "+str(j)+"\n")
            data.append(imgpath)
            label.append(j)
        label_list.append([category_name,j])
        j = j + 1

    allstr = ''.join(all_list)
    f = open('all_list.txt','w',encoding='utf-8')
    f.write(allstr)    return all_list , i , data , label, label_list
all_list,all_lenth,data,label,label_list = get_all_txt()print(data[0])print(label[0])print(label_list)

# 图片总数量查看print(all_lenth)print(len(data))print(len(label))

from sklearn.model_selection import train_test_split

train_img, val_img, train_label, val_label = train_test_split(data, label, test_size=0.2)

print(train_img[0])print(train_label[0])print(val_img[0])print(val_label[0])

import paddle.vision.transforms as T

train_transforms = T.Compose([
    T.Resize((512,512)),#随机选取目标尺寸调整图像大小
    T.RandomCrop((500,500)),#随机裁剪图像
    T.RandomHorizontalFlip(prob=0.7),
    T.RandomVerticalFlip(prob=0.7),
    T.RandomRotation(90),
    T.Transpose(),    # HWC -> CHW
    T.Normalize(
        to_rgb=True)])#对图像进行标准化eval_transforms = T.Compose([
    T.Resize((500,500)),#以一定的概率对图像进行随机像素内容变换，可包括亮度、对比度、饱和度、色相角度、通道顺序的调整
    T.Transpose(),    # HWC -> CHW
    T.Normalize(
        to_rgb=True)#对图像进行标准化])

## 构建dataset类import paddlefrom paddle.io import Datasetfrom PIL import Imageimport numpy as npclass MyDataset(Dataset):
    def __init__(self, mode = 'train'):
        # 训练样本数量
        self.training_data, self.training_label, self.test_data, self.test_label = train_img, train_label, val_img, val_label        if mode  == 'train':
            self.num_samples = len(train_img)        else:
            self.num_samples = len(val_img)
        self.mode = mode    def __getitem__(self, idx):
        if self.mode == 'train':
            image = self.training_data[idx]
            label = self.training_label[idx]
            img = Image.open(image).convert('RGB')
            img = train_transforms(img).astype("float32")        else:
            image = self.test_data[idx]
            label = self.test_label[idx]
            img = Image.open(image).convert('RGB')
            img = eval_transforms(img).astype("float32")        
        return img, np.array(label, dtype='int64')    def __len__(self):
        # 返回样本总数量
        return self.num_samples

# 训练的数据提供器train_dataset = MyDataset(mode='train')# 测试的数据提供器eval_dataset = MyDataset(mode='val')# 查看训练和测试数据的大小print('train大小：', len(train_dataset))print('eval大小：', len(eval_dataset))# 查看图片数据、大小及标签for data, label in train_dataset:    print(np.array(data).shape)    print(label)    break

import paddlefrom paddle.vision.models import MobileNetV2

model = MobileNetV2(num_classes=20)

paddle.summary(model, (1,3,500,500))  # 模型结构查看

import paddlefrom paddle.vision.models import MobileNetV2

model = MobileNetV2(num_classes=20)
model = paddle.Model(model)

EPOCH_NUM = 250BATCH_SIZE = 32learning_rate = 2e-3scheduler = paddle.optimizer.SGD(learning_rate=learning_rate, parameters=model.parameters())# 配置优化器、损失函数、评估指标model.prepare(scheduler, 
              paddle.nn.CrossEntropyLoss(), 
              paddle.metric.Accuracy())# 训练可视化VisualDL工具的回调函数visualdl = paddle.callbacks.VisualDL(log_dir='visualdl_log')# 启动模型全流程训练model.fit(train_dataset,
    eval_dataset,
    epochs=EPOCH_NUM,
    batch_size=BATCH_SIZE,
    callbacks=[visualdl],
    save_dir='output',
    shuffle=False,
    save_freq=10)

Epoch 250/250step 10/28 - loss: 0.2091 - acc: 0.9812 - 767ms/stepstep 20/28 - loss: 0.2418 - acc: 0.9812 - 768ms/stepstep 28/28 - loss: 0.1428 - acc: 0.9784 - 754ms/step

# 加载模型model = MobileNetV2(num_classes=20)
para_dict = paddle.load('output/final.pdparams') # 这里我根据上面的图中我选择使用220轮的模型进行验证model.load_dict(para_dict)

image ='work/动物1-20/貂熊/貂熊13.jpg'img = Image.open(image).convert('RGB')
img = paddle.to_tensor(train_transforms(img),dtype='float32')
img = paddle.reshape(img,(1,3,500,500))
pre = model(img)
result = pre.numpy().argmax()print(result)print(label_list[result+1]) # 加1 是因为前面生成的label_list第一个是 ['', -1]