最近一直在擼貓，為了貓主子的事情憂三愁四甲献，皺紋多了不少，頭發(fā)也掉了好幾根颂翼，神態(tài)也多了幾分憂郁晃洒，唯一不變的還是那份閑魚的懶散和浪蕩的心。

要說到深度學(xué)習圖像分類的經(jīng)典案例之一朦乏，那就是貓狗大戰(zhàn)了球及。貓和狗在外觀上的差別還是挺明顯的，無論是體型呻疹、四肢吃引、臉龐和毛發(fā)等等， 都是能通過肉眼很容易區(qū)分的刽锤。

那么如何讓機器來識別貓和狗呢镊尺？這就需要使用卷積神經(jīng)網(wǎng)絡(luò)來實現(xiàn)了。

網(wǎng)上已經(jīng)有不少人寫過這案例了并思，我也來嘗試下練練手庐氮。

完整代碼：https://github.com/ADlead/Dogs-Cats.git

ps:本文使用的是tensorflow1.x的版本，tf2.0版本可看下面這篇：
深度學(xué)習-python貓狗識別tensorflow2.0

一. 數(shù)據(jù)集的準備和預(yù)處理

貓狗照片的數(shù)據(jù)集直接從kaggle官網(wǎng)(https://www.kaggle.com/c/dogs-vs-cats)下載即可纺荧，下載后解壓旭愧，可以看到有訓(xùn)練集和測試集

train文件夾

test1文件夾

深度學(xué)習的框架使用的是tensorflow桃熄，為了能讓tensorflow分批輸入數(shù)據(jù)進行訓(xùn)練，我把所有的圖像像素信息存儲成batch文件型奥。訓(xùn)練集100個batch文件瞳收，每個文件有200張圖像。測試集1個batch文件厢汹，共5000張圖像螟深。

存儲成batch的代碼如下：

import cv2 as cv
import os
import numpy as np

import random
import pickle

import time

start_time = time.time()

data_dir = './data'
batch_save_path = './batch_files'

# 創(chuàng)建batch文件存儲的文件夾
os.makedirs(batch_save_path, exist_ok=True)

# 圖片統(tǒng)一大小：100 * 100
# 訓(xùn)練集 20000：100個batch文件烫葬，每個文件200張圖片
# 驗證集 5000：一個測試文件界弧，測試時 50張 x 100 批次

# 進入圖片數(shù)據(jù)的目錄凡蜻，讀取圖片信息
all_data_files = os.listdir(os.path.join(data_dir, 'train/'))

# print(all_data_files)

# 打算數(shù)據(jù)的順序
random.shuffle(all_data_files)

all_train_files = all_data_files[:20000]
all_test_files = all_data_files[20000:]

train_data = []
train_label = []
train_filenames = []

test_data = []
test_label = []
test_filenames = []

# 訓(xùn)練集
for each in all_train_files:
    img = cv.imread(os.path.join(data_dir,'train/',each),1)
    resized_img = cv.resize(img, (100,100))

    img_data = np.array(resized_img)
    train_data.append(img_data)
    if 'cat' in each:
        train_label.append(0)
    elif 'dog' in each:
        train_label.append(1)
    else:
        raise Exception('%s is wrong train file'%(each))
    train_filenames.append(each)

# 測試集
for each in all_test_files:
    img = cv.imread(os.path.join(data_dir,'train/',each), 1)
    resized_img = cv.resize(img, (100,100))

    img_data = np.array(resized_img)
    test_data.append(img_data)
    if 'cat' in each:
        test_label.append(0)
    elif 'dog' in each:
        test_label.append(1)
    else:
        raise Exception('%s is wrong test file'%(each))
    test_filenames.append(each)

print(len(train_data), len(test_data))

# 制作100個batch文件
start = 0
end = 200
for num in range(1, 101):
    batch_data = train_data[start: end]
    batch_label = train_label[start: end]
    batch_filenames = train_filenames[start: end]
    batch_name = 'training batch {} of 15'.format(num)

    all_data = {
    'data':batch_data,
    'label':batch_label,
    'filenames':batch_filenames,
    'name':batch_name
    }

    with open(os.path.join(batch_save_path, 'train_batch_{}'.format(num)), 'wb') as f:
        pickle.dump(all_data, f)

    start += 200
    end += 200

# 制作測試文件
all_test_data = {
    'data':test_data,
    'label':test_label,
    'filenames':test_filenames,
    'name':'test batch 1 of 1'
    }

with open(os.path.join(batch_save_path, 'test_batch'), 'wb') as f:
    pickle.dump(all_test_data, f)

end_time = time.time()
print('制作結(jié)束, 用時{}秒'.format(end_time - start_time))

運行程序后，文件就處理好了

二. 神經(jīng)網(wǎng)絡(luò)的編寫

cnn卷積神經(jīng)網(wǎng)絡(luò)的編寫如下垢箕，編寫卷積層划栓、池化層和全連接層的代碼

conv1_1 = tf.layers.conv2d(x, 16, (3, 3), padding='same', activation=tf.nn.relu, name='conv1_1')
conv1_2 = tf.layers.conv2d(conv1_1, 16, (3, 3), padding='same', activation=tf.nn.relu, name='conv1_2')
pool1 = tf.layers.max_pooling2d(conv1_2, (2, 2), (2, 2), name='pool1')
conv2_1 = tf.layers.conv2d(pool1, 32, (3, 3), padding='same', activation=tf.nn.relu, name='conv2_1')
conv2_2 = tf.layers.conv2d(conv2_1, 32, (3, 3), padding='same', activation=tf.nn.relu, name='conv2_2')
pool2 = tf.layers.max_pooling2d(conv2_2, (2, 2), (2, 2), name='pool2')
conv3_1 = tf.layers.conv2d(pool2, 64, (3, 3), padding='same', activation=tf.nn.relu, name='conv3_1')
conv3_2 = tf.layers.conv2d(conv3_1, 64, (3, 3), padding='same', activation=tf.nn.relu, name='conv3_2')
pool3 = tf.layers.max_pooling2d(conv3_2, (2, 2), (2, 2), name='pool3')
conv4_1 = tf.layers.conv2d(pool3, 128, (3, 3), padding='same', activation=tf.nn.relu, name='conv4_1')
conv4_2 = tf.layers.conv2d(conv4_1, 128, (3, 3), padding='same', activation=tf.nn.relu, name='conv4_2')
pool4 = tf.layers.max_pooling2d(conv4_2, (2, 2), (2, 2), name='pool4')

flatten = tf.layers.flatten(pool4)
fc1 = tf.layers.dense(flatten, 512, tf.nn.relu)
fc1_dropout = tf.nn.dropout(fc1, keep_prob=keep_prob)
fc2 = tf.layers.dense(fc1, 256, tf.nn.relu)
fc2_dropout = tf.nn.dropout(fc2, keep_prob=keep_prob)
fc3 = tf.layers.dense(fc2, 2, None)

三. Tensorflow計算圖的構(gòu)建

然后，再搭建tensorflow的計算圖条获，定義占位符忠荞，計算損失函數(shù)、預(yù)測值和準確率等等

self.x = tf.placeholder(tf.float32, [None, IMAGE_SIZE, IMAGE_SIZE, 3], 'input_data')
self.y = tf.placeholder(tf.int64, [None], 'output_data')
self.keep_prob = tf.placeholder(tf.float32)

# 圖片輸入網(wǎng)絡(luò)中
fc = self.conv_net(self.x, self.keep_prob)
self.loss = tf.losses.sparse_softmax_cross_entropy(labels=self.y, logits=fc)
self.y_ = tf.nn.softmax(fc) # 計算每一類的概率
self.predict = tf.argmax(fc, 1)
self.acc = tf.reduce_mean(tf.cast(tf.equal(self.predict, self.y), tf.float32))
self.train_op = tf.train.AdamOptimizer(LEARNING_RATE).minimize(self.loss)
self.saver = tf.train.Saver(max_to_keep=1)

最后的saver是要將訓(xùn)練好的模型保存到本地帅掘。

四. 模型的訓(xùn)練和測試

然后編寫訓(xùn)練部分的代碼委煤，訓(xùn)練步驟為1萬步

acc_list = []
with tf.Session() as sess:
    sess.run(tf.global_variables_initializer())

    for i in range(TRAIN_STEP):
        train_data, train_label, _ = self.batch_train_data.next_batch(TRAIN_SIZE)

        eval_ops = [self.loss, self.acc, self.train_op]
        eval_ops_results = sess.run(eval_ops, feed_dict={
            self.x:train_data,
            self.y:train_label,
            self.keep_prob:0.7
        })
        loss_val, train_acc = eval_ops_results[0:2]

        acc_list.append(train_acc)
        if (i+1) % 100 == 0:
            acc_mean = np.mean(acc_list)
            print('step:{0},loss:{1:.5},acc:{2:.5},acc_mean:{3:.5}'.format(
                i+1,loss_val,train_acc,acc_mean
            ))
        if (i+1) % 1000 == 0:
            test_acc_list = []
            for j in range(TEST_STEP):
                test_data, test_label, _ = self.batch_test_data.next_batch(TRAIN_SIZE)
                acc_val = sess.run([self.acc],feed_dict={
                    self.x:test_data,
                    self.y:test_label,
                    self.keep_prob:1.0
            })
            test_acc_list.append(acc_val)
            print('[Test ] step:{0}, mean_acc:{1:.5}'.format(
                i+1, np.mean(test_acc_list)
            ))
    # 保存訓(xùn)練后的模型
    os.makedirs(SAVE_PATH, exist_ok=True)
    self.saver.save(sess, SAVE_PATH + 'my_model.ckpt')

訓(xùn)練結(jié)果如下

訓(xùn)練1萬步后模型測試的平均準確率有0.82。

五. 識別和分類

最后修档，使用自己訓(xùn)練好的模型素标，把官網(wǎng)的測試圖片(共12500張)識別后進行分類，并將分類后的圖片分別寫入到兩個文件夾中萍悴，結(jié)果如下

分類為狗的圖片

分類為貓的圖片

可以看出头遭，分類之后還有少數(shù)分類不對的結(jié)果⊙⒂眨看來模型還有待提升计维，還可以調(diào)整網(wǎng)絡(luò)參數(shù)、調(diào)整學(xué)習率和學(xué)習步數(shù)撕予、使用圖像增強等技術(shù)對模型識別準確率進行提高鲫惶。

搞了一番，手和腳也有些累了实抡。是時候擼擼貓繼續(xù)閑魚了欠母。。

-----------------------------分割線------------------------------------

另外又寫了一篇tensorflow的圖像分類筆記吆寨，歡迎觀摩：
http://www.reibang.com/p/6bf4657ccd8e