这两天换了一种方法，是VGG方法，加了一些Dropout层，效果也是80%的样子。精度仍然还需要提高，下面直接附上我的代码：
import tensorflow as tf import pandas as pd import numpy as np import os #选择硬件设备# os.environ['CUDA_VISIBLE_DEVICES'] = '1' #选择硬件设备# #用pandas读取csv格式的训练集和测试集# emnist_train = pd.read_csv('emnist-csv/emnist-balanced-train.csv', header=None) emnist_test = pd.read_csv('emnist-csv/emnist-balanced-test.csv', header=None) #用pandas读取csv格式的训练集和测试集# #切割训练集# nd1 = emnist_train.iloc[:,1:] #切出除第一列以外的数据 label_train = emnist_train.iloc[:,0] #切出第一列数据作为标签 print("训练集标签的形状：", label_train.shape,'\n' "训练集标签的总数：", label_train.size,'\n' "训练集标签的维度：", label_train.ndim,'\n') #用于查看标签信息 label_train = pd.get_dummies(label_train) #将测试集标签进行one-hot编码 print("训练集的形状：",label_train.shape) #查看one-hot处理后标签的形状 assert label_train.shape[0] == label_train.shape[0] #确认训练集和训练集标签第一列是否同形状 label_train = label_train.values #将训练集放入数组存储 x_train = nd1.values #将训练集标签放入数组存储 #切割训练集# print('这里分割训练集和测试集') #切割测试集# nd2 = emnist_test.iloc[:,1:] #切出除第一列以外的数据 label_test = emnist_test.iloc[:,0] #切出第一列数据作为标签 print("测试集标签的形状：",label_test.shape,'\n' "测试集标签的总数：",label_test.size,'\n' "测试集标签的维度：",label_test.ndim,'\n') #用于查看标签信息 label_test = pd.get_dummies(label_test) #将测试集标签进行one-hot编码 print("测试集的形状：",label_test.shape) #查看one-hot处理后标签的形状 assert label_test.shape[0] == label_test.shape[0] #确认测试集和测试集标签第一列是否同形状 label_test = label_test.values #将测试集放入数组存储 x_test = nd2.values #将测试集标签放入数组存储 #切割测试集# # np.save('E:/Project/HandwritingRead/NumpyData/label',label_train) # np.save('E:/Project/HandwritingRead/NumpyData/sample',x_train_reshape) # np.save('E:/Project/HandwritingRead/NumpyData/label_test',label_test) # np.save('E:/Project/HandwritingRead/NumpyData/sample_test',x_test_reshape) # x_v = np.load('NumpyData/sample.npy') # y_v = np.load('NumpyData/label.npy') # x_w = np.load('NumpyData/sample_test.npy') # y_w = np.load('NumpyData/label_test.npy') xs = tf.placeholder(tf.float32, [None, 784]) #取出一块？*（28*28）的区域用于存放数据 ys = tf.placeholder(tf.float32, [None, 47]) #取出一块？*47的区域用于存放标签 xs_cut = tf.reshape(xs, [-1, 28, 28, 1]) #将xs形状改变为四维数组，第一维自动计算 # xc_cut = tf.reshape(xc, [-1, 28, 28, 1]) #训练网络 xs_cut = tf.layers.batch_normalization(inputs=xs_cut) conv1 = tf.layers.conv2d(inputs=xs_cut, filters=32, kernel_size=(3, 3), strides=1, activation=tf.nn.relu, padding='same') conv1 = tf.layers.batch_normalization(inputs=conv1) conv2 = tf.layers.conv2d(inputs=conv1, filters=32, kernel_size=(3,3), strides=1, padding='same', activation=tf.nn.relu) conv2 = tf.layers.batch_normalization(inputs=conv2) conv3 = tf.layers.conv2d(inputs=conv2, filters=32, kernel_size=(3,3), strides=1, padding='same', activation=tf.nn.relu) pool1 = tf.layers.max_pooling2d(inputs=conv3, pool_size=(2,2), strides=2, padding='same') pool1 = tf.layers.batch_normalization(pool1) conv4 = tf.layers.conv2d(inputs=pool1, filters=32, kernel_size=(3,3), strides=1, activation=tf.nn.relu, padding='same') conv4 = tf.layers.batch_normalization(inputs=conv4) conv5 = tf.layers.conv2d(inputs=conv4, filters=32, kernel_size=(3,3), strides=1, activation=tf.nn.relu, padding='same') conv5 = tf.layers.batch_normalization(inputs=conv5) conv6 = tf.layers.conv2d(inputs=conv5, filters=32, kernel_size=(3,3), strides=1, activation=tf.nn.relu, padding='same') pool2 = tf.layers.max_pooling2d(inputs=conv6, pool_size=(2,2), strides=2, padding='same') pool2 = tf.layers.batch_normalization(inputs=pool2) flat = tf.layers.flatten(pool2) re = tf.reshape(flat,[-1,7*7*32]) # flat = tf.layers.dense(inputs=re, # units=1024, # activation=tf.nn.relu) flat1 = tf.layers.dense(inputs=re, units=47) out = tf.nn.softmax(flat1) #训练网络# cross_entropy = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits_v2(labels = ys,logits=flat1)) #使用交叉熵设置损失函数 train = tf.train.MomentumOptimizer(1e-3,momentum=0.5).minimize(cross_entropy) #降低损失函数 correct = tf.equal(tf.argmax(flat1, 1), tf.argmax(ys, 1)) compute_accuracy = tf.reduce_mean(tf.cast(correct, tf.float32)) summary_op = tf.summary.merge_all() sess_config = tf.ConfigProto() sess_config.gpu_options.per_process_gpu_memory_fraction = 0.70 with tf.Session(config=sess_config) as sess: # #sess.run([tf.global_variables_initializer(), iterator.initializer], feed_dict={xs: x_v, ys: y_v}) # xs_batch,ys_batch = data_element # xs_batch_np = np.array(xs_batch,dtype = tf.int64) # ys_batch_np = np.array(ys_batch,dtype = tf.int64) # print(xs_batch_np.shape) # print(ys_batch_np.shape) # xs_batch_reshape = xs_batch_np.reshape([None,28,28]) # ys_batch_reshape = ys_batch_np.reshape([None,1]) # print(xs_batch_np.dtype) # print(ys_batch_np.dtype) # sess.run([tf.global_variables_initializer(), iterator.initializer], feed_dict={xs: xs_batch_np, ys: ys_batch_np}) # for i in range(1000): # x_w_batch, y_w_batch = sess.run(dataset_test_element) # if i % 50 == 0: # print(compute_accuracy(x_w_batch, y_w_batch)) sess.run(tf.global_variables_initializer()) #初始化变量 for e in range(50): for i in range(1000): x_train_b = x_train[i*100:(i+1)*100] #每次取训练集数据100个 y_train_b = label_train[i*100:(i+1)*100] #每次取训练集标签100个 sess.run(train,feed_dict={xs: x_train_b,ys: y_train_b}) #将100个数据和100个标签分别送入xs和ys if e%2 == 0: accuracy_test = sess.run(compute_accuracy,feed_dict = {xs:x_test,ys:label_test}) #计算精度 # accuracy = sess.run(compute_accuracy(x_test,label_test)) accuracy_train = sess.run(compute_accuracy,feed_dict = {xs:x_train_b,ys:y_train_b}) print('当前测试精度：',accuracy_test,'当前训练精度：',accuracy_train,'训练数据误差: ',sess.run(cross_entropy, feed_dict={xs: x_train_b, ys: y_train_b}),'当前批次第',i,'批','当前循环第',e,'轮') if accuracy_test > 0.95 : writer = tf.summary.FileWriter('E:/Project/HandwritingRead/SessGraph', sess.graph) writer.close() sess.close() # saver = tf.train.Saver() # save_path=saver.save(sess,'E:/Project/Model/emnist/emnist_save_net.ckpt')k # print('存储路径：',save_path)数据集以及AlexNet方法我在这个链接里写过，现在放出来：https://blog.csdn.net/Dangerous_Zombie/article/details/102773968 后续还会寻找一些方法提高精度。
