本文所使用的開源數(shù)據(jù)集（kaggle貓狗大戰(zhàn)）：
https://www.kaggle.com/c/dogs-vs-cats
國內(nèi)百度網(wǎng)盤下載地址：
https://pan.baidu.com/s/12ab32UNYI_6o4i9HyX-n8Q

利用本文代碼訓(xùn)練并生成的模型（對應(yīng)項(xiàng)目中的model文件夾）：
https://pan.baidu.com/s/1tBkVQKoH2Fc_HsBxtdAxnQ

簡單介紹：
（需要預(yù)先安裝pip install opencv-python, pip install flask, pip install tensorflow/pip install tensorflow-gpu）
本文使用Python3娃善，TensorFlow實(shí)現(xiàn)適合新手的VGG16模型（不了解VGG16的同學(xué)可以自行百度一下，本文沒有使用slim或者keras實(shí)現(xiàn)误债，對VGG16逐層實(shí)現(xiàn)，便于新手理解，有經(jīng)驗(yàn)的同學(xué)可以用高級庫重寫這部分）可應(yīng)用于單標(biāo)簽分類（一張圖片要么是貓遣钳，要么是狗）任務(wù)饱岸。

（預(yù)告：之后會寫一篇多標(biāo)簽分類任務(wù)驻龟，與單標(biāo)簽分類有些區(qū)別http://www.reibang.com/p/596db72a7e00）

整體訓(xùn)練邏輯：
0撤摸，使用pipeline方式異步讀取訓(xùn)練集圖片，節(jié)省內(nèi)存消耗褒纲，提高效率
1准夷，將圖像傳入到CNN（VGG16）中提取特征
2，將特征圖拉伸輸入到FC layer中得出分類預(yù)測向量
3莺掠，通過softmax交叉熵函數(shù)對預(yù)測向量和標(biāo)簽向量進(jìn)行訓(xùn)練衫嵌，得出最終模型
整體預(yù)測邏輯：
1，將圖像傳入到CNN（VGG16）中提取特征
2彻秆，將特征圖拉伸輸入到FC layer中得出分類預(yù)測向量
3楔绞，將預(yù)測向量做softmax操作，取向量中的最大值唇兑，并映射到對應(yīng)類別中
制作成web服務(wù)：
利用flask框架將整個(gè)項(xiàng)目啟動(dòng)成web服務(wù)酒朵，使得項(xiàng)目支持http方式調(diào)用
啟動(dòng)服務(wù)后調(diào)用以下地址測試
http://127.0.0.1:5050/dogOrCat?img_path=./data/test1/1.jpg
http://127.0.0.1:5050/dogOrCat?img_path=./data/test1/5.jpg

后續(xù)優(yōu)化邏輯：
可以采用遷移學(xué)習(xí)，模型融合等方案進(jìn)一步提高acc
可以左右翻轉(zhuǎn)圖片扎附，將訓(xùn)練集翻倍

運(yùn)行命令：
對數(shù)據(jù)集進(jìn)行訓(xùn)練：python DogVsCat.py train
對新的圖片進(jìn)行測試：python DogVsCat.py test
啟動(dòng)成http服務(wù)：python DogVsCat.py start

項(xiàng)目整體目錄結(jié)構(gòu)：

項(xiàng)目結(jié)構(gòu)

model結(jié)構(gòu)：

model文件夾結(jié)構(gòu)

訓(xùn)練過程：

epoch-1

epoch-2

epoch-3

epoch-13

epoch-13-中段

epoch-13-結(jié)束

整體代碼如下：

# coding:utf-8

import tensorflow as tf
import os, sys, random
import numpy as np
import cv2

from flask import request
from flask import Flask
import json
app = Flask(__name__)


class DogVsCat:
    def __init__(self):
        # 可調(diào)參數(shù)
        self.save_epoch = 1  # 每相隔多少個(gè)epoch保存一次模型
        self.train_max_num = 25000  # 訓(xùn)練時(shí)讀取的最大圖片數(shù)目 0~25000之間蔫耽，內(nèi)存不足的可以調(diào)小
        self.epoch_max = 13  # 最大迭代epoch次數(shù)
        self.batch_size = 16  # 訓(xùn)練時(shí)每個(gè)批次參與訓(xùn)練的圖像數(shù)目，顯存不足的可以調(diào)小
        self.class_num = 2  # 分類數(shù)目留夜，貓狗共兩類
        self.val_num = 20 * self.batch_size  # 不能大于self.train_max_num  做驗(yàn)證集用
        self.lr = 1e-4  # 初始學(xué)習(xí)率

        # 無需修改參數(shù)
        self.x_val = []
        self.y_val = []
        self.x = None  # 每批次的圖像數(shù)據(jù)
        self.y = None  # 每批次的one-hot標(biāo)簽
        self.learning_rate = None  # 學(xué)習(xí)率
        self.sess = None  # 持久化的tf.session
        self.pred = None  # cnn網(wǎng)絡(luò)結(jié)構(gòu)的預(yù)測
        self.keep_drop = tf.placeholder(tf.float32)  # dropout比例


    def dogOrCat(self, img_path):
        """
        貓狗分類
        :param img_path:
        :return:
        """
        im = cv2.imread(img_path)
        im = cv2.resize(im, (224, 224))
        im = [im]
        im = np.array(im, dtype=np.float32)
        im -= 147
        output = self.sess.run(self.output, feed_dict={self.x: im, self.keep_drop: 1.})
        ret = output.tolist()[0]
        ret = 'It is a cat' if ret[0] <= ret[1] else 'It is a dog'
        return ret


    def test(self, img_path):
        """
        測試接口
        :param img_path:
        :return:
        """
        self.x = tf.placeholder(tf.float32, [None, 224, 224, 3])  # 輸入數(shù)據(jù)
        self.pred = self.CNN()
        self.output = tf.nn.softmax(self.pred)

        saver = tf.train.Saver()
        # tfconfig = tf.ConfigProto(allow_soft_placement=True)
        # tfconfig.gpu_options.per_process_gpu_memory_fraction = 0.3  # 占用顯存的比例
        # self.ses = tf.Session(config=tfconfig)
        self.sess = tf.Session()
        self.sess.run(tf.global_variables_initializer())  # 全局tf變量初始化

        # 加載w,b參數(shù)
        saver.restore(self.sess, './model/DogVsCat-13')
        im = cv2.imread(img_path)
        im = cv2.resize(im, (224, 224))
        im = [im]
        im = np.array(im, dtype=np.float32)
        im -= 147
        output = self.sess.run(self.output, feed_dict={self.x: im, self.keep_drop: 1.})
        ret = output.tolist()[0]
        ret = 'It is a cat' if ret[0] <= ret[1] else 'It is a dog'
        print(ret)


    def train(self):
        """
        開始訓(xùn)練
        :return: 
        """
        self.x = tf.placeholder(tf.float32, [None, 224, 224, 3])  # 輸入數(shù)據(jù)
        self.y = tf.placeholder(tf.float32, [None, self.class_num])  # 標(biāo)簽數(shù)據(jù)
        self.learning_rate = tf.placeholder(tf.float32)  # 學(xué)習(xí)率

        # 生成訓(xùn)練用數(shù)據(jù)集
        x_train_list, y_train_list, x_val_list, y_val_list = self.getTrainDataset()

        print('開始轉(zhuǎn)換tensor隊(duì)列')
        x_train_list_tensor = tf.convert_to_tensor(x_train_list, dtype=tf.string)
        y_train_list_tensor = tf.convert_to_tensor(y_train_list, dtype=tf.float32)

        x_val_list_tensor = tf.convert_to_tensor(x_val_list, dtype=tf.string)
        y_val_list_tensor = tf.convert_to_tensor(y_val_list, dtype=tf.float32)

        x_train_queue = tf.train.slice_input_producer(tensor_list=[x_train_list_tensor], shuffle=False)
        y_train_queue = tf.train.slice_input_producer(tensor_list=[y_train_list_tensor], shuffle=False)

        x_val_queue = tf.train.slice_input_producer(tensor_list=[x_val_list_tensor], shuffle=False)
        y_val_queue = tf.train.slice_input_producer(tensor_list=[y_val_list_tensor], shuffle=False)

        train_im, train_label = self.dataset_opt(x_train_queue, y_train_queue)
        train_batch = tf.train.batch(tensors=[train_im, train_label], batch_size=self.batch_size, num_threads=2)

        val_im, val_label = self.dataset_opt(x_val_queue, y_val_queue)
        val_batch = tf.train.batch(tensors=[val_im, val_label], batch_size=self.batch_size, num_threads=2)


        # VGG16網(wǎng)絡(luò)
        print('開始加載網(wǎng)絡(luò)')
        self.pred = self.CNN()

        # 損失函數(shù)
        self.loss = tf.nn.softmax_cross_entropy_with_logits(logits=self.pred, labels=self.y)

        # 優(yōu)化器
        self.opt = tf.train.AdamOptimizer(learning_rate=self.learning_rate).minimize(self.loss)

        # acc
        self.acc_tf = tf.equal(tf.argmax(self.pred, 1), tf.argmax(self.y, 1))
        self.acc = tf.reduce_mean(tf.cast(self.acc_tf, tf.float32))

        with tf.Session() as self.sess:
            # 全局tf變量初始化
            self.sess.run(tf.global_variables_initializer())
            coordinator = tf.train.Coordinator()
            threads = tf.train.start_queue_runners(sess=self.sess, coord=coordinator)

            # 模型保存
            saver = tf.train.Saver()

            batch_max = len(x_train_list) // self.batch_size
            total_step = 1
            for epoch_num in range(self.epoch_max):
                lr = self.lr * (1 - (epoch_num/self.epoch_max) ** 2)  # 動(dòng)態(tài)學(xué)習(xí)率
                for batch_num in range(batch_max):
                    x_train_tmp, y_train_tmp = self.sess.run(train_batch)
                    self.sess.run(self.opt, feed_dict={self.x: x_train_tmp, self.y: y_train_tmp, self.learning_rate: lr, self.keep_drop: 0.5})

                    # 輸出評價(jià)標(biāo)準(zhǔn)
                    if total_step % 20 == 0 or total_step == 1:
                        print()
                        print('epoch:%d/%d batch:%d/%d step:%d lr:%.10f' % ((epoch_num + 1), self.epoch_max, (batch_num + 1), batch_max, total_step, lr))

                        # 輸出訓(xùn)練集評價(jià)
                        train_loss, train_acc = self.sess.run([self.loss, self.acc], feed_dict={self.x: x_train_tmp, self.y: y_train_tmp, self.keep_drop: 1.})
                        print('train_loss:%.10f  train_acc:%.10f' % (np.mean(train_loss), train_acc))

                        # 輸出驗(yàn)證集評價(jià)
                        val_loss_list, val_acc_list = [], []
                        for i in range(int(self.val_num/self.batch_size)):
                            x_val_tmp, y_val_tmp = self.sess.run(val_batch)
                            val_loss, val_acc = self.sess.run([self.loss, self.acc], feed_dict={self.x: x_val_tmp, self.y: y_val_tmp, self.keep_drop: 1.})
                            val_loss_list.append(np.mean(val_loss))
                            val_acc_list.append(np.mean(val_acc))
                        print('  val_loss:%.10f    val_acc:%.10f' % (np.mean(val_loss), np.mean(val_acc)))

                    total_step += 1

                # 保存模型
                if (epoch_num + 1) % self.save_epoch == 0:
                    print('正在保存模型：')
                    saver.save(self.sess, './model/DogVsCat', global_step=(epoch_num + 1))
            coordinator.request_stop()
            coordinator.join(threads)


    def CNN(self):
        """
        VGG16 + FC
        :return: 
        """
        # 權(quán)重
        weight = {
            # 輸入 batch_size*224*224*3

            # 第一層
            'wc1_1': tf.get_variable('wc1_1', [3, 3, 3, 64]),  # 卷積 輸出：batch_size*224*224*64
            'wc1_2': tf.get_variable('wc1_2', [3, 3, 64, 64]),  # 卷積 輸出：batch_size*224*224*64
            # 池化 輸出：112*112*64

            # 第二層
            'wc2_1': tf.get_variable('wc2_1', [3, 3, 64, 128]),  # 卷積 輸出：batch_size*112*112*128
            'wc2_2': tf.get_variable('wc2_2', [3, 3, 128, 128]),  # 卷積 輸出：batch_size*112*112*128
            # 池化 輸出：56*56*128

            # 第三層
            'wc3_1': tf.get_variable('wc3_1', [3, 3, 128, 256]),  # 卷積 輸出：batch_size*56*56*256
            'wc3_2': tf.get_variable('wc3_2', [3, 3, 256, 256]),  # 卷積 輸出：batch_size*56*56*256
            'wc3_3': tf.get_variable('wc3_3', [3, 3, 256, 256]),  # 卷積 輸出：batch_size*56*56*256
            # 池化 輸出：28*28*256

            # 第四層
            'wc4_1': tf.get_variable('wc4_1', [3, 3, 256, 512]),  # 卷積 輸出：batch_size*28*28*512
            'wc4_2': tf.get_variable('wc4_2', [3, 3, 512, 512]),  # 卷積 輸出：batch_size*28*28*512
            'wc4_3': tf.get_variable('wc4_3', [3, 3, 512, 512]),  # 卷積 輸出：batch_size*28*28*512
            # 池化 輸出：14*14*512

            # 第五層
            'wc5_1': tf.get_variable('wc5_1', [3, 3, 512, 512]),  # 卷積 輸出：batch_size*14*14*512
            'wc5_2': tf.get_variable('wc5_2', [3, 3, 512, 512]),  # 卷積 輸出：batch_size*14*14*512
            'wc5_3': tf.get_variable('wc5_3', [3, 3, 512, 512]),  # 卷積 輸出：batch_size*14*14*512
            # 池化 輸出：7*7*512

            # 全鏈接第一層
            'wfc_1': tf.get_variable('wfc_1', [7*7*512, 4096]),

            # 全鏈接第二層
            'wfc_2': tf.get_variable('wfc_2', [4096, 4096]),

            # 全鏈接第三層
            'wfc_3': tf.get_variable('wfc_3', [4096, self.class_num]),
        }

        # 偏移量
        biase = {
            # 第一層
            'bc1_1': tf.get_variable('bc1_1', [64]),
            'bc1_2': tf.get_variable('bc1_2', [64]),

            # 第二層
            'bc2_1': tf.get_variable('bc2_1', [128]),
            'bc2_2': tf.get_variable('bc2_2', [128]),

            # 第三層
            'bc3_1': tf.get_variable('bc3_1', [256]),
            'bc3_2': tf.get_variable('bc3_2', [256]),
            'bc3_3': tf.get_variable('bc3_3', [256]),

            # 第四層
            'bc4_1': tf.get_variable('bc4_1', [512]),
            'bc4_2': tf.get_variable('bc4_2', [512]),
            'bc4_3': tf.get_variable('bc4_3', [512]),

            # 第五層
            'bc5_1': tf.get_variable('bc5_1', [512]),
            'bc5_2': tf.get_variable('bc5_2', [512]),
            'bc5_3': tf.get_variable('bc5_3', [512]),

            # 全鏈接第一層
            'bfc_1': tf.get_variable('bfc_1', [4096]),

            # 全鏈接第二層
            'bfc_2': tf.get_variable('bfc_2', [4096]),

            # 全鏈接第三層
            'bfc_3': tf.get_variable('bfc_3', [self.class_num]),
        }

        # 第一層
        net = tf.nn.conv2d(input=self.x, filter=weight['wc1_1'], strides=[1, 1, 1, 1], padding='SAME')  # 卷積
        net = tf.nn.leaky_relu(tf.nn.bias_add(net, biase['bc1_1']))  # 加b 然后 激活
        net = tf.nn.conv2d(net, filter=weight['wc1_2'], strides=[1, 1, 1, 1], padding='SAME')  # 卷積
        net = tf.nn.leaky_relu(tf.nn.bias_add(net, biase['bc1_2']))  # 加b 然后 激活
        net = tf.nn.max_pool(value=net, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='VALID')  # 池化

        # 第二層
        net = tf.nn.conv2d(net, weight['wc2_1'], [1, 1, 1, 1], padding='SAME')  # 卷積
        net = tf.nn.leaky_relu(tf.nn.bias_add(net, biase['bc2_1']))  # 加b 然后 激活
        net = tf.nn.conv2d(net, weight['wc2_2'], [1, 1, 1, 1], padding='SAME')  # 卷積
        net = tf.nn.leaky_relu(tf.nn.bias_add(net, biase['bc2_2']))  # 加b 然后 激活
        net = tf.nn.max_pool(net, [1, 2, 2, 1], [1, 2, 2, 1], padding='VALID')  # 池化

        # 第三層
        net = tf.nn.conv2d(net, weight['wc3_1'], [1, 1, 1, 1], padding='SAME')  # 卷積
        net = tf.nn.leaky_relu(tf.nn.bias_add(net, biase['bc3_1']))  # 加b 然后 激活
        net = tf.nn.conv2d(net, weight['wc3_2'], [1, 1, 1, 1], padding='SAME')  # 卷積
        net = tf.nn.leaky_relu(tf.nn.bias_add(net, biase['bc3_2']))  # 加b 然后 激活
        net = tf.nn.conv2d(net, weight['wc3_3'], [1, 1, 1, 1], padding='SAME')  # 卷積
        net = tf.nn.leaky_relu(tf.nn.bias_add(net, biase['bc3_3']))  # 加b 然后 激活
        net = tf.nn.max_pool(net, [1, 2, 2, 1], [1, 2, 2, 1], padding='VALID')  # 池化

        # 第四層
        net = tf.nn.conv2d(net, weight['wc4_1'], [1, 1, 1, 1], padding='SAME')  # 卷積
        net = tf.nn.leaky_relu(tf.nn.bias_add(net, biase['bc4_1']))  # 加b 然后 激活
        net = tf.nn.conv2d(net, weight['wc4_2'], [1, 1, 1, 1], padding='SAME')  # 卷積
        net = tf.nn.leaky_relu(tf.nn.bias_add(net, biase['bc4_2']))  # 加b 然后 激活
        net = tf.nn.conv2d(net, weight['wc4_3'], [1, 1, 1, 1], padding='SAME')  # 卷積
        net = tf.nn.leaky_relu(tf.nn.bias_add(net, biase['bc4_3']))  # 加b 然后 激活
        net = tf.nn.max_pool(net, [1, 2, 2, 1], [1, 2, 2, 1], padding='VALID')  # 池化

        # 第五層
        net = tf.nn.conv2d(net, weight['wc5_1'], [1, 1, 1, 1], padding='SAME')  # 卷積
        net = tf.nn.leaky_relu(tf.nn.bias_add(net, biase['bc5_1']))  # 加b 然后 激活
        net = tf.nn.conv2d(net, weight['wc5_2'], [1, 1, 1, 1], padding='SAME')  # 卷積
        net = tf.nn.leaky_relu(tf.nn.bias_add(net, biase['bc5_2']))  # 加b 然后 激活
        net = tf.nn.conv2d(net, weight['wc5_3'], [1, 1, 1, 1], padding='SAME')  # 卷積
        net = tf.nn.leaky_relu(tf.nn.bias_add(net, biase['bc5_3']))  # 加b 然后 激活
        net = tf.nn.max_pool(net, [1, 2, 2, 1], [1, 2, 2, 1], padding='VALID')  # 池化
        print('last-net', net)

        # 拉伸flatten匙铡，把多個(gè)圖片同時(shí)分別拉伸成一條向量
        net = tf.reshape(net, shape=[-1, weight['wfc_1'].get_shape()[0]])
        print(weight['wfc_1'].get_shape()[0])
        print('拉伸flatten', net)

        # 全鏈接層
        # fc第一層
        net = tf.matmul(net, weight['wfc_1']) + biase['bfc_1']
        net = tf.nn.dropout(net, self.keep_drop)
        net = tf.nn.relu(net)
        print('fc第一層', net)

        # fc第二層
        net = tf.matmul(net, weight['wfc_2']) + biase['bfc_2']
        net = tf.nn.dropout(net, self.keep_drop)
        net = tf.nn.relu(net)
        print('fc第二層', net)

        # fc第三層
        net = tf.matmul(net, weight['wfc_3']) + biase['bfc_3']
        print('fc第三層', net)
        return net


    def getTrainDataset(self):
        """
        整理數(shù)據(jù)集，把圖像resize為224*224*3碍粥，訓(xùn)練集做成25000*224*224*3鳖眼，把label做成one-hot形式
        :return: 
        """
        train_data_list = os.listdir('./data/train_data/')
        print('共有%d張訓(xùn)練圖片， 讀取%d張：' % (len(train_data_list), self.train_max_num))
        random.shuffle(train_data_list)  # 打亂順序

        x_val_list = train_data_list[:self.val_num]
        y_val_list = [[0, 1] if file_name.find('cat') > -1 else [1, 0] for file_name in x_val_list]

        x_train_list = train_data_list[self.val_num:self.train_max_num]
        y_train_list = [[0, 1] if file_name.find('cat') > -1 else [1, 0] for file_name in x_train_list]

        return x_train_list, y_train_list, x_val_list, y_val_list


    def dataset_opt(self, x_train_queue, y_train_queue):
        """
        處理圖片和標(biāo)簽
        :param queue: 
        :return: 
        """
        queue = x_train_queue[0]
        contents = tf.read_file('./data/train_data/' + queue)
        im = tf.image.decode_jpeg(contents)
        im = tf.image.resize_images(images=im, size=[224, 224])
        im = tf.reshape(im, tf.stack([224, 224, 3]))
        im -= 147  # 去均值化
        # im /= 255  # 將像素處理在0~1之間嚼摩，加速收斂
        # im -= 0.5  # 將像素處理在-0.5~0.5之間
        return im, y_train_queue[0]



if __name__ == '__main__':
    opt_type = sys.argv[1:][0]

    instance = DogVsCat()

    if opt_type == 'train':
        instance.train()
    elif opt_type == 'test':
        instance.test('./data/test1/1.jpg')
    elif opt_type == 'start':
        # 將session持久化到內(nèi)存中
        instance.test('./data/test1/1.jpg')
        # 啟動(dòng)web服務(wù)
        # http://127.0.0.1:5050/dogOrCat?img_path=./data/test1/1.jpg
        @app.route('/dogOrCat', methods=['GET', 'POST'])
        def dogOrCat():
            img_path = ''
            if request.method == 'POST':
                img_path = request.form.to_dict().get('img_path')
            elif request.method == 'GET':
                # img_path = request.args.get('img_path')
                img_path = request.args.to_dict().get('img_path')
            print(img_path)
            ret = instance.dogOrCat(img_path)
            print(ret)
            return json.dumps({'type': ret})
        app.run(host='0.0.0.0', port=5050, debug=False)