一、用CNN实现手写数字识别
import tensorflow as tf import numpy as np from sklearn.datasets import load_digits import time print( time.ctime() ) digits = load_digits() X_data = digits.data.astype(np.float32) Y_data = digits.target.astype(np.float32).reshape(-1, 1) def generatebatch(X, Y, n_examples, batch_size): for batch_i in range(n_examples // batch_size): start = batch_i * batch_size end = start + batch_size batch_xs = X[start:end] batch_ys = Y[start:end] yield batch_xs, batch_ys # 生成每一个batch from sklearn.preprocessing import MinMaxScaler scaler = MinMaxScaler() X_data = scaler.fit_transform(X_data) from sklearn.preprocessing import OneHotEncoder Y = OneHotEncoder().fit_transform(Y_data).todense()#todense返回矩阵 #转换为图片格式(batch, height, width, channels) X = X_data.reshape(-1, 8, 8, 1) batch_size = 8 #使用MBGD算法,batch_size=8 tf.reset_default_graph()#tf.reset_default_graph函数用于清除默认图形堆栈并重置全局默认图形。 #输入层 tf_X = tf.placeholder(tf.float32, [None, 8, 8, 1]) tf_Y = tf.placeholder(tf.float32, [None, 10]) #卷积层,创建权重变量和偏置变量,其中的四维矩阵代表了filters的深度,即10个filter,其中每个filter的深度是1,若是RGB则每个filter深度为3 conv_filter_w1 = tf.Variable( tf.random_normal([3, 3, 1, 10]) )#3*3的filter,每个filter深度是1,共10个filter #conv_filter_w1 = tf.get_variable( 'weights', [3, 3, 1, 10], initializer=tf.truncated_normal_initializer(stddev=0.1) ) conv_filter_b1 = tf.Variable( tf.random_normal([10]) )#10个filter对应10个bias #conv_filter_b1 = tf.get_variable( 'biases', [10], initializer=tf.constant_initializer(0.1) ) #tf.nn.conv2d提供了方便的函数来实现卷积层的前向传播算法 #函数的第一个输入为当前层的节点矩阵,注意这个是四维矩阵即tf_X #其中(tf_X)后面三维对应一个节点矩阵,第一维对应一个输入batch,比如在输入层 #input[0, :, :, :]表示第一张图,input[1, :, :, :]表示第二张图,input[2, :, :, :]表示第三张图,以此类推 #tf.nn.conv2d第二个参数提供了卷积层权重,第三个参数是不同维度上的步长,长度为4的数组,第一维和最后一维数字要求一定是1 #最后一个参数是padding,SAME表示添加全0填充,VALID表示不添加 conv = tf.nn.conv2d( tf_X, conv_filter_w1, strides=[1, 1, 1, 1], padding='SAME') #tf.nn.bias_add提供了一个方便的函数给每个节点加上偏置项 #注意这里不能直接使用加法,因为矩阵上不同位置上的节点都需要加上同样的偏置项,也就是下面第二种直接加法是错误的 bias = tf.nn.bias_add( conv, conv_filter_b1 ) #bias = conv + conv_filter_b1这是错误的做法 #ReLU激活函数去线性化 relu_feature_maps1 = tf.nn.relu( bias ) print(relu_feature_maps1) #Tensor("Relu:0", shape=(?, 8, 8, 10), dtype=float32) # 池化层,tf.nn.max_pool实现了最大池化层的前向传播过程,参数和tf.nn.conv2d函数类似 #ksize提供了filter的尺寸,strides提供了步长,padding提供是否使用全0填充 max_pool1 = tf.nn.max_pool( relu_feature_maps1, ksize=[1,3,3,1], strides=[1,2,2,1], padding='SAME') print(max_pool1) #Tensor("MaxPool:0", shape=(?, 4,4,10), dtype=float32) # 卷积层2 conv_filter_w2 = tf.Variable( tf.random_normal([3, 3, 10, 5]) )#因为上一层的卷积层深度是10,即有10个输出,这里一个filter深度就是10,而用5个filter conv_filter_b2 = tf.Variable( tf.random_normal([5]) ) conv2 = tf.nn.conv2d(relu_feature_maps1, conv_filter_w2, strides=[1, 2, 2, 1], padding='SAME') conv_out2 = tf.nn.bias_add( conv2, conv_filter_b2 ) print(conv_out2) #Tensor("BiasAdd_1:0", shape=(?, 4, 4, 5), dtype=float32) # BN归一化层+激活层 #批标准化(batch normalization,BN)一般用在激活函数之前,使结果各个维度均值为0,方差为1。 #通过规范化让激活函数分布在线性区间,让每一层的输入有一个稳定的分布会有利于网络的训练。 batch_mean, batch_var = tf.nn.moments( conv_out2, axes=[0, 1, 2], keep_dims=True ) #axes:表示在哪个维度上求解,是个list,例如 [0, 1, 2] #keep_dims:是否保持维度 shift = tf.Variable(tf.zeros([5])) scale = tf.Variable(tf.ones([5])) epsilon = 1e-3 #•tf.nn.batch_normalization(x, mean, variance, offset, scale, variance_epsilon, name=None) #https://blog.csdn.net/lanchunhui/article/details/70792458 BN_out = tf.nn.batch_normalization( conv_out2, batch_mean, batch_var, shift, scale, epsilon ) print(BN_out) #Tensor("batchnorm/add_1:0", shape=(?, 4, 4, 5), dtype=float32) relu_BN_maps2 = tf.nn.relu(BN_out) # 池化层 max_pool2 = tf.nn.max_pool( relu_BN_maps2, ksize=[1,3,3,1], strides=[1,2,2,1], padding='SAME' ) print(max_pool2) #Tensor("MaxPool_1:0", shape=(?, 2, 2, 5), dtype=float32) # 将特征图进行展开 max_pool2_flat = tf.reshape( max_pool2, [-1, 2*2*5] ) # 全连接层 fc_w1 = tf.Variable( tf.random_normal([2*2*5,50]) ) fc_b1 = tf.Variable( tf.random_normal([50]) ) fc_out1 = tf.nn.relu( tf.matmul(max_pool2_flat, fc_w1) + fc_b1 ) # 输出层 out_w1 = tf.Variable( tf.random_normal([50,10]) ) out_b1 = tf.Variable( tf.random_normal([10]) ) pred = tf.nn.softmax( tf.matmul(fc_out1, out_w1) + out_b1 ) loss = -tf.reduce_mean( tf_Y * tf.log( tf.clip_by_value(pred,1e-11,1.0) ) ) train_step = tf.train.AdamOptimizer(1e-3).minimize(loss) y_pred = tf.arg_max(pred,1) bool_pred = tf.equal( tf.arg_max(tf_Y,1), y_pred ) accuracy = tf.reduce_mean( tf.cast(bool_pred,tf.float32) ) # 准确率 with tf.Session() as sess: sess.run(tf.global_variables_initializer()) for epoch in range(1000): # 迭代1000个周期 for batch_xs, batch_ys in generatebatch(X, Y, Y.shape[0], batch_size): # 每个周期进行MBGD算法 sess.run( train_step, feed_dict={tf_X:batch_xs, tf_Y:batch_ys} ) if(epoch%100 == 0): res = sess.run(accuracy,feed_dict={ tf_X:X,tf_Y:Y} ) print (epoch, res) res_ypred = y_pred.eval(feed_dict={tf_X:X,tf_Y:Y}).flatten() # 只能预测一批样本,不能预测一个样本 print (res_ypred) from sklearn.metrics import accuracy_score print ( accuracy_score( Y_data,res_ypred.reshape(-1,1) )) print( time.ctime() )