#!/usr/bin/env python # -*- coding: utf-8 -*- # @Time : 2018/3/30 上午8:26 # @Author : ruina.sun # @Software: PyCharm Community Edition import tensorflow as tf import numpy as np import matplotlib.pyplot as plt from tensorflow.examples.tutorials.mnist import input_data import os pro_path = os.path.abspath('..') # 样本个数是位置权值参数个数的5-30倍 # receptive field 感受野，猫看球视觉实验，部分视觉中枢的神经元是亮的，-->局部感受野。 # convolution （卷积）-->划窗，点乘求和 # pooling （池化）-->划窗，取max/mean值 # padding （填充)卷积和池化都存在窗格不够的情况，都需要padding处理方式。 # 步长为(1,1)的话：same-padding保留边缘信息，可能会补零以得到和原来平面一样大小， # valid-padding有可能缺失边缘信息，不补零可能会得到比原来平面小的平面。 '''多通道计算: https://blog.csdn.net/yudiemiaomiao/article/details/72466402''' mnist = input_data.read_data_sets('MNIST_data', one_hot=True) batch_size = 100 n_batch = mnist.train.num_examples // batch_size # 截尾正太初始化权值 def weight_variable(shape): initial = tf.truncated_normal(shape, stddev=0.1) return tf.Variable(initial) # 偏置初始化为0.1 def bias_variable(shape): # initial = tf.constant(shape, 0.1) initial = tf.constant(0.1, shape=shape) return tf.Variable(initial) # 卷积层 def conv2d(x, W): '''参数W就是卷积窗口''' '''x: data_format: "NHWC":[batch, in_height, in_width, in_channels] ''' '''W: filter / kernel tensor of shape, 卷积窗大小: [filter_height, filter_width, in_channels, out_channels]''' '''W: 卷积窗大小: [卷积窗高, 卷积窗宽, 输入图层个数, 单位输入图层的输出图层个数或称卷积窗/核个数(有几个卷积窗/核，每一个输入图层就有几个输出图层)]''' '''stride: stride[0] = stride[3] = 1, stride[1]代表x方向步长, stride[1]代表y方向步长''' '''padding: 'SAME' or 'VALID' ''' return tf.nn.conv2d(x, W, strides=[1, 1, 1, 1], padding='SAME') # 池化层 def max_pooling_2x2(x): '''池化没有任何参数，直接画切窗口区域，取最大取均值''' '''ksize[0] = stride[3] = 1, 池化窗大小：ksize[1]代表窗长/高?, ksize[1]代表窗高/长？''' '''stride[0] = stride[3] = 1, stride[1]代表x方向步长, stride[2]代表y方向步长''' return tf.nn.max_pool(x, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='SAME') print('*******', mnist.test.labels.shape) # 定义两个placeholder x = tf.placeholder(tf.float32, [None, 784]) # 28*28 y = tf.placeholder(tf.float32, [None, 10]) # 改变x的格式转为4D的向量[batch, in_height, in_width, in_channels] x_image = tf.reshape(x, [-1, 28, 28, 1]) # -1之后会变成100，1表示是黑白的图片，rgb的话会是3,对于3通道图像的各通道而言，是在每个通道上分别执行二维卷积，然后将3个通道加起来，得到该位置的二维卷积输出。 # 初始化第一个卷积层的权值和偏置, each filter is a chanel W_conv1 = weight_variable([5, 5, 1, 32]) # 采用32个5*5*1的卷积核，从一个黑白输入图层抽取特征，得到32个特征平面。如果是rgb的话需要把1改成3。 b_conv1 = bias_variable([32]) # 每一个卷积核有一个偏置值/每一个输出面or体有一个偏置值？ # 把x_image和权值向量进行卷积，再加上偏置值，然后应用于relu激活函数 h_conv1 = tf.nn.relu(conv2d(x_image, W_conv1) + b_conv1) h_pool1 = max_pooling_2x2(h_conv1) # 初始化第二个卷积层的权值和偏置, each filter is a chanel W_conv2 = weight_variable([5, 5, 32, 64]) # 采用64个5*5*32的卷积核，从一个黑白输入图层抽取特征，得到64个特征平面。32个通道会对应相加，最终还是得到64个map b_conv2 = bias_variable([64]) # 每一个卷积核有一个偏置值/每一个输出面or体有一个偏置值？ # 把h_pool1和权值向量进行卷积，再加上偏置值，然后应用于relu激活函数 h_conv2 = tf.nn.relu(conv2d(h_pool1, W_conv2) + b_conv2) h_pool2 = max_pooling_2x2(h_conv2) # 28*28的图片经过第一次的卷积(SAME)后,得到28*28，池化后14*14*32 # 第一次的卷积(SAME)后,得到14*14，池化后7*7*64 # 经过上面的两层操作得到64张7*7的平面，或者说7*7*64的立方体。 # 把池化层2的输出扁平化为1维，拍平 h_pool2_flat = tf.reshape(h_pool2, [-1, 7 * 7 * 64]) # 初始化第一个全连接层的权值 W_fc1 = weight_variable([7 * 7 * 64, 1024]) # 上一层的输入有7*7*64个神经元，全连接到1024个神经元上。 b_fc1 = bias_variable([1024]) # 求第一个全连接层的输出 h_fc1 = tf.nn.relu(tf.matmul(h_pool2_flat, W_fc1) + b_fc1) # 对全连接的输出drop一下，降低过拟合 keep_prob = tf.placeholder(tf.float32) h_fc1_drop = tf.nn.dropout(h_fc1, keep_prob) # 初始化第二个全连接层 W_fc2 = weight_variable([1024, 10]) # 上一层的输入有7*7*64个神经元，全连接到1024个神经元上。 b_fc2 = bias_variable([10]) # 计算输出 prediction = tf.nn.softmax((tf.matmul(h_fc1_drop, W_fc2) + b_fc2)) # 交叉熵代价函数 cross_entropy = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(labels=y, logits=prediction)) # 使用优化器 train_step = tf.train.AdamOptimizer(learning_rate=1e-4).minimize(cross_entropy) # 结果存放在一个布尔列表中 correct_prediction = tf.equal(tf.argmax(prediction, 1), tf.argmax(y, 1)) # argmax返回一维张量中最大的值位置下标 # 求准确率 accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32)) with tf.Session() as sess: sess.run(tf.global_variables_initializer()) for epoch in range(21): for batch in range(n_batch): batch_xs, batch_ys = mnist.train.next_batch(batch_size) print('@@@@@@@', batch_ys.shape) sess.run(train_step, feed_dict={x: batch_xs, y: batch_ys, keep_prob: 0.7}) acc = sess.run(accuracy, feed_dict={x: mnist.test.images, y: mnist.test.labels, keep_prob: 1.0}) print('Iter' + str(epoch) + ', Testing Accuracy=' + str(acc)) # Iter0, Testing Accuracy=0.8552 # Iter1, Testing Accuracy=0.9664 # Iter2, Testing Accuracy=0.9785 # Iter3, Testing Accuracy=0.981 # Iter4, Testing Accuracy=0.9834 # Iter5, Testing Accuracy=0.9858 # Iter6, Testing Accuracy=0.9867 # Iter7, Testing Accuracy=0.9858 # Iter8, Testing Accuracy=0.9873 # Iter9, Testing Accuracy=0.9856 # Iter10, Testing Accuracy=0.9874 # Iter11, Testing Accuracy=0.9897 # Iter12, Testing Accuracy=0.9898 # Iter13, Testing Accuracy=0.9884 # Iter14, Testing Accuracy=0.9897 # Iter15, Testing Accuracy=0.9895 # Iter16, Testing Accuracy=0.9903 # Iter17, Testing Accuracy=0.9906 # Iter18, Testing Accuracy=0.9912 # Iter19, Testing Accuracy=0.9904 # Iter20, Testing Accuracy=0.9921
