去年研一的时候想做kaggle上的一道题目：猫狗分类，但是苦于对卷积神经网络一直没有很好的认识，现在把这篇文章的内容补上去。（部分代码参考网上的，我改变了卷积神经网络的网络结构，其实主要部分我加了一层1X1的卷积层，至于作用，我会在后文详细介绍）

题目地址：猫狗大战

同时数据集也可以在上面下载到。

既然是手把手，那么就要从前期的导入数据开始：

1. 导入数据
2. #import sys, io
3. #sys.stdout = io.TextIOWrapper(sys.stdout.buffer,encoding='utf8') # Change default encoding to utf8
4. #coding=utf-8
5. import tensorflow as tf
6. import numpy as np
7. import os
8. train_dir='G:/data/CatVSdogtrain/train/' #训练数据的文件夹，这里你们要换成自己的
9. file_dir=train_dir
10. #定义一个函数把训练样本和测试样本集合起来
11. def get_files(file_dir):
12. '''''
13. input:
14. 训练图片放的图片集
15. returns：
16. 图片列表和标签列表
17. '''
18. cats=[]
19. label_cats=[]
20. dogs=[]
21. label_dogs=[]
22. # file=[]
23. for file in os.listdir(file_dir):
24. # file=np.hstack[files,file]
25. name=file.split(sep='.')
26. # print (name)
27. if name[0]=='cat':
28. cats.append(file_dir+file)
29. label_cats.append(0)
30. else:
31. dogs.append(file_dir+file)
32. label_dogs.append(1)
33. print ('there is %d cats and %d dogs' %(len(cats),len(dogs)))
34. #打乱文件的顺序，其实在获取batch的时候也可以做，但是为了方便还是在这里做了
35. image_list=np.hstack((cats,dogs))
36. label_list=np.hstack((label_cats,label_dogs))
37.  
38. temp=np.array([image_list,label_list])
39. temp=temp.transpose()
40. np.random.shuffle(temp)#打乱顺序函数
41. image_list=list(temp[:,0])
42. label_list=list(temp[:,1])
43. label_list = [int(i) for i in label_list]
44.  
45. return image_list,label_list

    其实这一段没什么好说的，无非就是做好训练样本，和标签。。。。代码仅供参考。

    2 get_batch

    def get_batch(image,label,image_W,image_H,batch_size,capacity):

    #image, label:生成的batch的图像和标签list

    #image_w, image_H:图片的大小

    #batch_size: 每个batch共有多少张图片

    #capacity ：队列的容量

    # return：图像和标签的batch

    # image=image_list

    # label=label_list

    #转换格式，让python 可以识别的格式,其实就是两个tensor

    image=tf.cast(image,tf.string)

    label=tf.cast(label,tf.int32)

    #生成队列

    input_queue=tf.train.slice_input_producer([image,label])

    label=input_queue[1]

    image_contents=tf.read_file(input_queue[0])

    image=tf.image.decode_jpeg(image_contents,channels=3)

    ##数据增强应该在这里

    image = tf.image.resize_image_with_crop_or_pad(image, image_W, image_H)

    image=tf.image.per_image_standardization(image)

    image_batch,label_batch=tf.train.batch([image,label],batch_size=batch_size,

    num_threads=64,capacity=capacity)

    label_batch=tf.reshape(label_batch,[batch_size])

    image_batch=tf.cast(image_batch,tf.float32)

    return image_batch,label_batch

    为什么要设置一个batch，一个batch 呢？

    如果损失函数是非凸的话，整个样本就算在超级计算机上可以算的动，也会卡在局部最优上，分批训练表示全样本的抽样实现，也就相当于人为引入修正梯度上的采样噪声，使'一路不通找别路'更有可能搜索最优值。

    其中LCLR 2017上有一篇文章专门讨论了这个问题：On Large-Batch Training for Deep Learning: Generalization Gap and Sharp Minima

    

    3建立卷积神经网络

    import tensorflow as tf

    def inference(images,batch_size,n_classes):

    ''''Build the model

    args:

    images:images batch, 4D tensor ,tf,float32,[batch_size,width,height,channels]

    returns:

    output tensor with the computed logits,floar, [batch_size,n_classes]

    #conv1,shape=[kernel size, kernel size,channels, kernel numbers]

    '''

    with tf.variable_scope('conv1') as scope:

    weights=tf.get_variable('weights',shape=[1,1,3, 16],dtype=tf.float32,

    initializer=tf.truncated_normal_initializer(stddev=0.1,dtype=tf.float32))

    biases=tf.get_variable('biases',shape=[16],dtype=tf.float32,

    initializer=tf.constant_initializer(0.1))

    conv=tf.nn.conv2d(images,weights,strides=[1,1,1,1],padding='SAME')

    pre_activation=tf.nn.bias_add(conv,biases)

    conv1=tf.nn.relu(pre_activation,name=scope.name)

    #poo11 and norm1

    with tf.variable_scope('pooling1_lrn') as scope:

    pool1=tf.nn.max_pool(conv1,ksize=[1,3,3,1],strides=[1,2,2,1],

    padding='SAME',name='pooling1')

    normal=tf.nn.lrn(pool1,depth_radius=4,bias=1.0,alpha=0.001/9.0,beta=0.75,name='norm1')

    #conv2

    with tf.variable_scope('conv2') as scope:

    weights=tf.get_variable('weights',shape=[3,3,16, 16],dtype=tf.float32,

    initializer=tf.truncated_normal_initializer(stddev=0.1,dtype=tf.float32))

    biases=tf.get_variable('biases',shape=[16],dtype=tf.float32,

    initializer=tf.constant_initializer(0.1))

    conv=tf.nn.conv2d(normal,weights,strides=[1,1,1,1],padding='SAME')

    pre_activation=tf.nn.bias_add(conv,biases)

    conv2=tf.nn.relu(pre_activation,name=scope.name)

    #pool2 and norm2

    with tf.variable_scope('pooling1_2rn') as scope:

    pool2=tf.nn.max_pool(conv2,ksize=[1,3,3,1],strides=[1,2,2,1],

    padding='SAME',name='pooling2')

    norma2=tf.nn.lrn(pool2,depth_radius=4,bias=1.0,alpha=0.001/9.0,beta=0.75,name='norm2')

    ##conv3

    with tf.variable_scope('conv3') as scope:

    weights=tf.get_variable('weights',shape=[3,3,16,16],dtype=tf.float32,

    initializer=tf.truncated_normal_initializer(stddev=0.1,dtype=tf.float32))

    biases=tf.get_variable('biases',shape=[16],dtype=tf.float32,

    initializer=tf.constant_initializer(0.1))

    conv=tf.nn.conv2d(norma2,weights,strides=[1,1,1,1],padding='SAME')

    pre_activation=tf.nn.bias_add(conv,biases)

    conv2=tf.nn.relu(pre_activation,name=scope.name)

    #poo11 and norm1

    with tf.variable_scope('pooling3_lrn') as scope:

    norma3=tf.nn.lrn(conv2,depth_radius=4,bias=1.0,alpha=0.001/9.0,beta=0.75,name='norm3')

    pool3=tf.nn.max_pool(norma3,ksize=[1,3,3,1],strides=[1,1,1,1],

    padding='SAME',name='pooling3')

    # # local3

    with tf.variable_scope('local3') as scope:

    reshape=tf.reshape(pool3,shape=[batch_size,-1])

    dim=reshape.get_shape()[1].value

    weights=tf.get_variable('weights',shape=[dim,128],dtype=tf.float32,

    initializer=tf.truncated_normal_initializer(stddev=0.005,dtype=tf.float32))

    biases=tf.get_variable('biases',shape=[128],dtype=tf.float32,

    initializer=tf.constant_initializer(0.1))

    local3=tf.nn.relu(tf.matmul(reshape,weights)+biases,name=scope.name)

    # #local4

    # with tf.variable_scope('local4') as scope:

    # weights = tf.get_variable('weights',

    # shape=[128,128],

    # dtype=tf.float32,

    # initializer=tf.truncated_normal_initializer(stddev=0.005,dtype=tf.float32))

    # biases = tf.get_variable('biases',

    # shape=[128],

    # dtype=tf.float32,

    # initializer=tf.constant_initializer(0.1))

    # local4 = tf.nn.relu(tf.matmul(local3, weights) + biases, name='local4')

    #local4

    with tf.variable_scope('local4') as scope:

    weights=tf.get_variable('weights',shape=[128,128],dtype=tf.float32,

    initializer=tf.truncated_normal_initializer(stddev=0.005,dtype=tf.float32))

    biases=tf.get_variable('biases',shape=[128],dtype=tf.float32,

    initializer=tf.constant_initializer(0.1))

    local4=tf.nn.relu(tf.matmul(local3,weights)+biases,name='local4')

    #softmax

    with tf.variable_scope('softmax_linear') as scope:

    weights=tf.get_variable('softmax_linear',shape=[128,n_classes],dtype=tf.float32,

    initializer=tf.truncated_normal_initializer(stddev=0.005,dtype=tf.float32))

    biases = tf.get_variable('biases',

    shape=[n_classes],

    dtype=tf.float32,

    initializer=tf.constant_initializer(0.1))

    softmax_linear=tf.add(tf.matmul(local4,weights),biases,name='softmax_linear')

    return softmax_linear

    这里面，我建立了一个1X1的卷积核，建立这个卷积核的作用主要有以下几个方面考虑：

    假设如果这个1X1卷积层的输入与输出都是一个平面，那么1X1卷积仅仅可以对数据进行非线性变化，但是它是完全不考虑像素与周边其他像素关系。但卷记得输入输出如果是长方体，所以1X1卷积实际上是对每个像素点在不同的channels上进行线性组合（信息整合），同时保留了图片原有的平面结构，通过调节depth，从而完成升维或者降维的功能。

    如下图，如果选择2个filters 的1X1 卷积层，那么数据就从原本的depth3 降到2.若用4个filters ,那么就起到了升维的作用。

    

    我的整个网络包括三个卷积层，三个全连接层。

    4损失函数部分

    def losses(logits,labels):

    with tf.variable_scope('loss') as scope:

    # cross_entropy=tf.nn.sparse_softmax_cross_entropy_with_logits\(logits=logits,labels=labels,name='xentropy_per_example')

    cross_entropy = tf.nn.sparse_softmax_cross_entropy_with_logits(logits=logits, labels=labels, name='xentropy_per_example')

    loss=tf.reduce_mean(cross_entropy,name='loss')

    tf.summary.scalar(scope.name+'/loss',loss)

    return loss

    def training(loss,learning_rate):

    with tf.name_scope('optimizer'):

    optimizer=tf.train.AdamOptimizer(learning_rate=learning_rate)

    global_step = tf.Variable(0, name='global_step', trainable=False)

    train_op=optimizer.minimize(loss,global_step=global_step)

    return train_op

    def evaluation(logits,labels):

    with tf.variable_scope('accuracy') as scope:

    correct=tf.nn.in_top_k(logits,labels,1)

    correct=tf.cast(correct,tf.float16)

    accuracy=tf.reduce_mean(correct)

    tf.summary.scalar(scope.name+'/accuracy',accuracy)

    return accuracy

    这部分没什么好讲的，从tensorflow官网上有相似的例程，就是按照那个编写的。损失函数就是最常用的softmax损失函数。优化方法是AdamOptimizer。。。感觉tensorflow最让我爽的点就是这里不用自己求梯度。。。曾经因为求梯度，头发掉了一地。。。。

    5training

    # -*- coding: utf-8 -*-

    """

    Created on Fri Oct 13 08:42:54 2017

    @author: Administrator

    """

    import os

    import numpy as np

    import tensorflow as tf

    import myinput_data

    import mymodel

    ##

    learning_rate=0.0001

    ##

    def run_training():

    train_dir='G:/data/CatVSdogtrain/train/'

    logs_train_dir='G:/data/CatVSdogtrain/logits/train/'

    train,train_label=myinput_data.get_files(train_dir)

    train_batch,train_label_batch=myinput_data.get_batch(train,train_label,

    IMAGE_W,

    IMAGE_H,

    BATCH_SIZE,

    CAPACITY

    )

    train_logits=mymodel.inference(train_batch,BATCH_SIZE,N_CLASSES)

    train_loss=mymodel.losses(train_logits,train_label_batch)

    train_op=mymodel.training(train_loss,learning_rate)

    train_acc=mymodel.evaluation(train_logits,train_label_batch)

    summary_op=tf.summary.merge_all()

    sess=tf.Session()

    train_writer=tf.summary.FileWriter(logs_train_dir,sess.graph)

    saver=tf.train.Saver()

    sess.run(tf.global_variables_initializer())

    coord=tf.train.Coordinator()

    threads=tf.train.start_queue_runners(sess=sess,coord=coord)

    try:

    for step in np.arange(MAX_STEP):

    if coord.should_stop():

    break

    _, tra_loss,tra_acc=sess.run([train_op,train_loss,train_acc])

    if step % 50==0:

    print ('Step %d,train loss=%.2f, train accuracy=%.2f%%'% (step,tra_loss,tra_acc*100.0))

    summary_str = sess.run(summary_op)

    train_writer.add_summary(summary_str, step)

    if step % 2000 == 0 or (step + 1) == MAX_STEP:

    checkpoint_path = os.path.join(logs_train_dir, 'model.ckpt')

    saver.save(sess, checkpoint_path, global_step=step)

    except tf.errors.OutOfRangeError:

    print('Done training -- epoch limit reached')

    finally:

    coord.request_stop()

    coord.join(threads)

    sess.close()

    这一部分就是保存训练结果，然后把损失函数调到最小。。。识别率就会高，编写可以参照tensorflow的例程。

    6 mytest

    from PIL import Image

    import tensorflow as tf

    import matplotlib.pyplot as plt

    import numpy as np

    import myinput_data

    import mytraining

    import mymodel

    def get_one_image(train):

    ##随机的选取一张图片

    ##return :ndarry

    n=len(train)

    ind=np.random.randint(0,n)

    img_dir=train[ind]

    image=Image.open(img_dir)

    plt.imshow(image)

    image=image.resize([208,208])

    image=np.array(image)

    return image

    def evaluate_one_image():

    train_dir='G:/data/CatVSdogtrain/train/'

    train,train_label=myinput_data.get_files(train_dir)

    image_array=get_one_image(train)

    with tf.Graph().as_default():

    image=tf.cast(image_array, tf.float32)

    image=tf.image.per_image_standardization(image)

    image=tf.reshape(image,[1,208,208,3])

    logit=mymodel.inference(image,BATCH_SIZE,N_CLASSES)

    x=tf.placeholder(tf.float32,shape=[208,208,3])

    logs_train_dir='G:/data/CatVSdogtrain/logits/train/'

    saver=tf.train.Saver()

    with tf.Session() as sess:

    print("Reading checkpoints...")

    ckpt = tf.train.get_checkpoint_state(logs_train_dir)

    if ckpt and ckpt.model_checkpoint_path:

    global_step = ckpt.model_checkpoint_path.split('/')[-1].split('-')[-1]

    saver.restore(sess, ckpt.model_checkpoint_path)

    print('Loading success, global_step is %s' % global_step)

    # print(sess.run())

    else:

    print('No checkpoint file found')

    prediction = sess.run(logit, feed_dict={x: image_array})

    max_index = np.argmax(prediction)

    if max_index==0:

    print('This is a cat with possibility %.6f' %prediction[:, 0])

    print('This is a dog with possibility %.6f' %prediction[:, 1])

    else:

    print('This is a dog with possibility %.6f' %prediction[:, 1])

    print('This is a cat with possibility %.6f' %prediction[:, 0])

    运行这一段代码，然后在命令行执行evaluate_one_image()

    结果如下：

    

    这个只是最简单的卷积神经网络，所以说整个实现过程很简单，但是追求远远不止这些，如果大家有什么对卷积的想法可以一起交流。