稍稍乱入的CNN,本文依然是学习周莫烦视频的笔记。
还有 google 在 udacity 上的 CNN 教程。
CNN(Convolutional Neural Networks) 卷积神经网络简单讲就是把一个图片的数据传递给CNN,原涂层是由RGB组成,然后CNN把它的厚度加厚,长宽变小,每做一层都这样被拉长,最后形成一个分类器:
![](http://upload-images.jianshu.io/upload_images/1667471-3d06daf10515c57f.png?imageMogr2/auto-orient/strip%7CimageView2/2/w/1240)
![](http://upload-images.jianshu.io/upload_images/1667471-82842bb1e4e5c927.png?imageMogr2/auto-orient/strip%7CimageView2/2/w/1240)
如果想要分成十类的话,那么就会有0到9这十个位置,这个数据属于哪一类就在哪个位置上是1,而在其它位置上为零。
在 RGB 这个层,每一次把一块核心抽出来,然后厚度加厚,长宽变小,形成分类器:
![](http://upload-images.jianshu.io/upload_images/1667471-52f2f5a4a6699bdf.png?imageMogr2/auto-orient/strip%7CimageView2/2/w/1240)
在 CNN 中有几个重要的概念:
- stride
- padding
- pooling
stride,就是每跨多少步抽取信息。每一块抽取一部分信息,长宽就缩减,但是厚度增加。抽取的各个小块儿,再把它们合并起来,就变成一个压缩后的立方体。
padding,抽取的方式有两种,一种是抽取后的长和宽缩减,另一种是抽取后的长和宽和原来的一样。
pooling,就是当跨步比较大的时候,它会漏掉一些重要的信息,为了解决这样的问题,就加上一层叫pooling,事先把这些必要的信息存储起来,然后再变成压缩后的层:
![](http://upload-images.jianshu.io/upload_images/1667471-7fda4b9d7141db3c.png?imageMogr2/auto-orient/strip%7CimageView2/2/w/1240)
patch, 就是小方块的长宽的像素,in size 是image的厚度为1,out size是输出的厚度为32:
![](http://upload-images.jianshu.io/upload_images/1667471-a1c239ca2c056a28.png?imageMogr2/auto-orient/strip%7CimageView2/2/w/1240)
CNN的结构,分析一张图片时,先放一个CNN的图层,再把这个图层进行一个pooling。这样可以比较好的保持信息,之后再加第二层的CNN和pooling。
导入一个图片之后,先是有它的RGB三个图层,然后把像素块缩小变厚。本来有三个厚度,然后把它变成八个厚度,它的长宽在不断的减小,最后把它们连接在一起:
![](http://upload-images.jianshu.io/upload_images/1667471-0d067b4d78aeb0d1.png?imageMogr2/auto-orient/strip%7CimageView2/2/w/1240)
下面就是用 tensorflow 构建一个 CNN 的代码,
里面主要有4个layer,分别是:
- convolutional layer1 + max pooling;
- convolutional layer2 + max pooling;
- fully connected layer1 + dropout;
- fully connected layer2 to prediction.
import tensorflow as tf
from tensorflow.examples.tutorials.mnist import input_data # number 1 to 10 data mnist = input_data.read_data_sets('MNIST_data', one_hot=True) def compute_accuracy(v_xs, v_ys): global prediction y_pre = sess.run(prediction, feed_dict={xs: v_xs, keep_prob: 1}) correct_prediction = tf.equal(tf.argmax(y_pre,1), tf.argmax(v_ys,1)) accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32)) result = sess.run(accuracy, feed_dict={xs: v_xs, ys: v_ys, keep_prob: 1}) return result # 产生随机变量,符合 normal 分布 # 传递 shape 就可以返回weight和bias的变量 def weight_variable(shape): initial = tf.truncated_normal(shape, stddev=0.1) return tf.Variable(initial) def bias_variable(shape): initial = tf.constant(0.1, shape=shape) return tf.Variable(initial) # 定义2维的 convolutional 图层 def conv2d(x, W): # stride [1, x_movement, y_movement, 1] # Must have strides[0] = strides[3] = 1 # strides 就是跨多大步抽取信息 return tf.nn.conv2d(x, W, strides=[1, 1, 1, 1], padding='SAME') # 定义 pooling 图层 def max_pool_2x2(x): # stride [1, x_movement, y_movement, 1] # 用pooling对付跨步大丢失信息问题 return tf.nn.max_pool(x, ksize=[1,2,2,1], strides=[1,2,2,1], padding='SAME') # define placeholder for inputs to network xs = tf.placeholder(tf.float32, [None, 784]) # 784=28x28 ys = tf.placeholder(tf.float32, [None, 10]) keep_prob = tf.placeholder(tf.float32) x_image = tf.reshape(xs, [-1, 28, 28, 1]) # 最后一个1表示数据是黑白的 # print(x_image.shape) # [n_samples, 28,28,1] ## 1. conv1 layer ## # 把x_image的厚度1加厚变成了32 W_conv1 = weight_variable([5, 5, 1, 32]) # patch 5x5, in size 1, out size 32 b_conv1 = bias_variable([32]) # 构建第一个convolutional层,外面再加一个非线性化的处理relu h_conv1 = tf.nn.relu(conv2d(x_image, W_conv1) + b_conv1) # output size 28x28x32 # 经过pooling后,长宽缩小为14x14 h_pool1 = max_pool_2x2(h_conv1) # output size 14x14x32 ## 2. conv2 layer ## # 把厚度32加厚变成了64 W_conv2 = weight_variable([5,5, 32, 64]) # patch 5x5, in size 32, out size 64 b_conv2 = bias_variable([64]) # 构建第二个convolutional层 h_conv2 = tf.nn.relu(conv2d(h_pool1, W_conv2) + b_conv2) # output size 14x14x64 # 经过pooling后,长宽缩小为7x7 h_pool2 = max_pool_2x2(h_conv2) # output size 7x7x64 ## 3. func1 layer ## # 飞的更高变成1024 W_fc1 = weight_variable([7*7*64, 1024]) b_fc1 = bias_variable([1024]) # [n_samples, 7, 7, 64] ->> [n_samples, 7*7*64] # 把pooling后的结果变平 h_pool2_flat = tf.reshape(h_pool2, [-1, 7*7*64]) h_fc1 = tf.nn.relu(tf.matmul(h_pool2_flat, W_fc1) + b_fc1) h_fc1_drop = tf.nn.dropout(h_fc1, keep_prob) ## 4. func2 layer ## # 最后一层,输入1024,输出size 10,用 softmax 计算概率进行分类的处理 W_fc2 = weight_variable([1024, 10]) b_fc2 = bias_variable([10]) prediction = tf.nn.softmax(tf.matmul(h_fc1_drop, W_fc2) + b_fc2) # the error between prediction and real data cross_entropy = tf.reduce_mean(-tf.reduce_sum(ys * tf.log(prediction), reduction_indices=[1])) # loss train_step = tf.train.AdamOptimizer(1e-4).minimize(cross_entropy) sess = tf.Session() # important step sess.run(tf.initialize_all_variables()) for i in range(1000): batch_xs, batch_ys = mnist.train.next_batch(100) sess.run(train_step, feed_dict={xs: batch_xs, ys: batch_ys, keep_prob: 0.5}) if i % 50 == 0: print(compute_accuracy( mnist.test.images, mnist.test.labels))
转自:http://www.jianshu.com/p/e2f62043d02b