【AI达人创造营】菜市场常见鱼类分类从0到PaddleLite部署

# 解压缩数据!unzip -oq -d images data/data103322/images.zip

# 数据加载和预处理import osimport paddleimport numpy as npimport paddlehub.vision.transforms as T# 定义数据集class FishDataset(paddle.io.Dataset):
    def __init__(self, dataset_dir, transforms, mode='train'):
        # 数据集存放路径
        self.dataset_dir = dataset_dir        # 数据增强
        self.transforms = transforms        # 分类数
        self.label_lst = []
        self.num_classes= self.get_label()
        self.mode = mode        # 根据mode读取对应的数据集
        if self.mode == 'train':
            self.file = 'train_list.txt'
        elif self.mode == 'test':
            self.file = 'test_list.txt'
        else:
            self.file = 'validate_list.txt'
        self.file = os.path.join(self.dataset_dir, self.file)        with open(self.file, 'r') as f:
            self.data = f.read().split('\n')[:-1]    
    def get_label(self):
        # 获取分类数
        with open(os.path.join(dataset_dir, 'label_list.txt'), 'r') as f:
            labels = f.readlines()        for idx, label in enumerate(labels):
            dic = {}
            dic['label_name'] = label.split('\n')[0]
            dic['label_id'] = idx
            self.label_lst.append(dic)        return len(self.label_lst)    def __getitem__(self, idx):
        img_path, label = self.data[idx].split(' ')
        img_path = os.path.join(self.dataset_dir, img_path)
        im = self.transforms(img_path)        return im, int(label)    
    def __len__(self):
        return len(self.data)# 定义数据增强train_Transforms = T.Compose([
                T.Resize((256, 256)),
                T.CenterCrop(224),
                T.RandomHorizontalFlip(),
                T.Normalize()
                ], to_rgb=True)
eval_Transforms = T.Compose([
                T.Resize((256, 256)),
                T.CenterCrop(224),
                T.Normalize()
                ], to_rgb=True)# 读取数据集dataset_dir = 'images/images'fish_train = FishDataset(dataset_dir, train_Transforms)
fish_validate = FishDataset(dataset_dir, eval_Transforms, mode='validate')print('训练集的图片数量: {}'.format(len(fish_train)))print('验证集的图片数量: {}'.format(len(fish_validate)))print('分类数: {}'.format(len(fish_train.label_lst)))

# label_id转labelid2label = {}for i in fish_train.label_lst:
     id2label[i['label_id']] = i['label_name']print(id2label)

# 鱼类英文名转中文en2zh = {'Cuttlefish': '墨鱼', 'Turbot': '多宝鱼', 'Hairtail': '带鱼', 'Grouper': '石斑鱼', 'Saury': '秋刀鱼', 'Octopus': '章鱼', 'Red_fish': '红鱼', 'Tilapia_mossambica': '罗非鱼', 'Variegated_carp': '胖头鱼', 'Grass_Carp': '草鱼', 'Silverfish': '银鱼', 'Herring': '青鱼', 'Horsehead_fish': '马头鱼', 'Squid': '鱿鱼', 'Catfish': '鲇鱼', 'Perch': '鲈鱼', 'Abalone': '鲍鱼', 'Salmon': '鲑鱼', 'Silver_carp': '鲢鱼', 'Carp': '鲤鱼', 'Crucian_carp': '鲫鱼', 'Silvery_pomfret': '鲳鱼', 'Bream': '鲷鱼', 'Plaice': '鲽鱼', 'Parabramis_pekinensis': '鳊鱼', 'Eel': '鳗鱼', 'Yellow_croaker': '黄鱼', 'Ricefield_eel': '黄鳝', 'Snakehead': '黑鱼', 'Bibcock_fish': '龙头鱼'}print(en2zh)

# 生成fish_label_list.txt文件with open('fish_label_list.txt', 'w') as f:    for i, value in enumerate(en2zh.values()):
        f.write('{} {}\n'.format(i, value))

from PIL import Imageimport matplotlib.pyplot as plt

path = 'images/images/train'plt.figure(figsize=(30, 8))for idx, name in enumerate(en2zh.keys()):    for fpath, dirname, fname in os.walk(os.path.join(path, name)):
        plt.subplot(3, 10, idx+1)
        img = Image.open(os.path.join(fpath, fname[0]))
        plt.title(name)
        plt.imshow(img)

import paddlefrom paddle.vision.models import mobilenet_v2# 设置pretrained参数为True，可以加载mobilenet_v2在imagenet数据集上的预训练模型model = paddle.Model(mobilenet_v2(pretrained=True, num_classes=len(fish_train.label_lst)))

from paddle.optimizer import Momentumfrom paddle.regularizer import L2Decayfrom paddle.nn import CrossEntropyLossfrom paddle.metric import Accuracy# 配置优化器optimizer = Momentum(learning_rate=0.001,
                     momentum=0.9,
                     weight_decay=L2Decay(1e-4),
                     parameters=model.parameters())# 进行训练前准备model.prepare(optimizer, CrossEntropyLoss(), Accuracy(topk=(1, 5)))# 启动训练model.fit(fish_train,
          fish_validate,
          epochs=35,
          batch_size=8,
          save_dir="./output")

if not os.path.exists('final'):
    os.mkdir('final')# 将final.pdparams复制到final文件夹!cp output/final.pdparams final

# 模型评估，根据prepare接口配置的loss和metric进行返回result = model.evaluate(fish_validate)print(result)

# 批量预测from paddle.static import InputSpec# 加载final模型参数inputs = InputSpec([None, 1*1*3*224*224], 'float32', 'x')
labels = InputSpec([None, 30], 'int32', 'x')
model = paddle.Model(mobilenet_v2(num_classes=len(fish_train.label_lst)), inputs, labels)# 加载模型参数model.load('final/final.pdparams')# 定义优化器optimizer = Momentum(learning_rate=0.001,
                     momentum=0.9,
                     weight_decay=L2Decay(1e-4),
                     parameters=model.parameters())# 进行预测前准备model.prepare(optimizer, CrossEntropyLoss(), Accuracy(topk=(1, 5)))# 加载测试集数据fish_test = FishDataset(dataset_dir, eval_Transforms, mode='test')# 进行预测操作result = model.predict(fish_test)# 定义画图方法def show_img(idx, predict):
    with open(os.path.join(dataset_dir, 'test_list.txt')) as f:
        data = f.readlines()
    plt.figure()    print('predict: {}'.format(predict))
    img = Image.open(os.path.join(dataset_dir, data[idx].split()[0]))
    plt.imshow(img)
    plt.show()# 抽样展示indexs = [2, 15, 38, 100]for idx in indexs:
    show_img(idx, en2zh[id2label[np.argmax(result[0][idx])]])

# 单张图片预测# 读取单张图片image = paddle.to_tensor(fish_test[80][0]).reshape([1, 1, 3, 224, 224])
image_id = 80# 单张图片预测result = model.predict(image)# 可视化结果show_img(image_id, en2zh[id2label[np.argmax(result)]])

# save inference modelfrom paddle.static import InputSpecfrom paddle.vision.models import mobilenet_v2import paddle
inputs = InputSpec([None, 1*1*3*224*224], 'float32', 'x')
labels = InputSpec([None, 30], 'int32', 'x')
paddle.disable_static()
model = mobilenet_v2(num_classes=30)# 加载训练好的模型参数state_dict = paddle.load("final/final.pdparams")# 将训练好的参数读取到网络中model.set_state_dict(state_dict)# 设置模型为评估模式model.eval()# 保存inference模型paddle.jit.save(
    layer=model,
    path="inference/fish_predict",
    input_spec=[InputSpec(shape=[None, 3, 224, 224], dtype='float32')])    
print("==>Inference model saved in inference.")

import numpy as npimport paddleimport paddle.nn.functional as F# 确保从paddle.vision.datasets.MNIST中加载的图像数据是np.ndarray类型paddle.vision.set_image_backend('cv2')# 将测试图像数据转换为tensor，并reshape为[1, 784]test_image = paddle.reshape(paddle.to_tensor(fish_test[100][0]), [1, 3, 224, 224])print('true label: ', fish_test[100][1])# 然后执行图像归一化# 加载保存的模型loaded_model = paddle.jit.load("./inference/fish_predict")# 利用加载的模型执行预测preds = loaded_model(test_image)
pred_label = paddle.argmax(preds)# 打印预测结果print("The predicted label is : ", pred_label.numpy())

!pip install paddlelite

# 准备PaddleLite部署模型#--valid_targets中参数（arm）用于传统手机，（npu,arm ）用于华为带有npu处理器的手机!paddle_lite_opt --model_file=inference/fish_predict.pdmodel \
      --param_file=inference/fish_predict.pdiparams \
      --valid_targets=arm \
      --optimize_out_type=naive_buffer \
      --optimize_out=paddleLite_fish_predict_opt    #   --valid_targets=npu,arm

    <string name="MODEL_PATH_DEFAULT">models/fish_predict</string>
    <string name="LABEL_PATH_DEFAULT">labels/fish_label_list.txt</string>
    <string name="IMAGE_PATH_DEFAULT">images/predict_demo.jpg</string>