# 部分函数请参考前一篇或后一篇文章
import tensorflow as tf
import tfrecords2array
import numpy as np
import matplotlib.pyplot as plt
from collections import OrderedDict
def lenet(char_classes):
y_train = []
x_train = []
y_test = []
x_test = []
for char_class in char_classes:
train_data = tfrecords2array.tfrecord2array(
r"./data_tfrecords/" + char_class + "_tfrecords/train.tfrecords")
test_data = tfrecords2array.tfrecord2array(
r"./data_tfrecords/" + char_class + "_tfrecords/test.tfrecords")
y_train.append(train_data[0])
x_train.append(train_data[1])
y_test.append(test_data[0])
x_test.append(test_data[1])
for i in [y_train, x_train, y_test, x_test]:
for j in i:
print(j.shape)
y_train = np.vstack(y_train)
x_train = np.vstack(x_train)
y_test = np.vstack(y_test)
x_test = np.vstack(x_test)
class_num = y_test.shape[-1]
print("x_train.shape=" + str(x_train.shape))
print("x_test.shape=" + str(x_test.shape))
sess = tf.InteractiveSession()
x = tf.placeholder("float", shape=[None, 784])
y_ = tf.placeholder("float", shape=[None, class_num])
# 把x更改为4维张量，第1维代表样本数量，第2维和第3维代表图像长宽， 第4维代表图像通道数, 1表示黑白
x_image = tf.reshape(x, [-1, 28, 28, 1])
# 第一层：卷积层
conv1_weights = tf.get_variable(
"conv1_weights",
[5, 5, 1, 32],
initializer=tf.truncated_normal_initializer(stddev=0.1))
# 过滤器大小为5*5, 当前层深度为1， 过滤器的深度为32
conv1_biases = tf.get_variable("conv1_biases", [32],
initializer=tf.constant_initializer(0.0))
conv1 = tf.nn.conv2d(x_image, conv1_weights, strides=[1, 1, 1, 1],
padding='SAME')
# 移动步长为1, 使用全0填充
relu1 = tf.nn.relu(tf.nn.bias_add(conv1, conv1_biases)) # 激活函数Relu去线性化
# 第二层：最大池化层
# 池化层过滤器的大小为2*2, 移动步长为2，使用全0填充
pool1 = tf.nn.max_pool(relu1, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1],
padding='SAME')
# 第三层：卷积层
conv2_weights = tf.get_variable(
"conv2_weights",
[5, 5, 32, 64],
initializer=tf.truncated_normal_initializer(stddev=0.1))
# 过滤器大小为5*5, 当前层深度为32， 过滤器的深度为64
conv2_biases = tf.get_variable(
"conv2_biases", [64], initializer=tf.constant_initializer(0.0))
conv2 = tf.nn.conv2d(pool1, conv2_weights, strides=[1, 1, 1, 1],
padding='SAME')
# 移动步长为1, 使用全0填充
relu2 = tf.nn.relu(tf.nn.bias_add(conv2, conv2_biases))
# 第四层：最大池化层
# 池化层过滤器的大小为2*2, 移动步长为2，使用全0填充
pool2 = tf.nn.max_pool(relu2, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1],
padding='SAME')
# 第五层：全连接层
fc1_weights = tf.get_variable("fc1_weights", [7 * 7 * 64, 1024],
initializer=tf.truncated_normal_initializer(
stddev=0.1))
# 7*7*64=3136把前一层的输出变成特征向量
fc1_biases = tf.get_variable(
"fc1_biases", [1024], initializer=tf.constant_initializer(0.1))
pool2_vector = tf.reshape(pool2, [-1, 7 * 7 * 64])
fc1 = tf.nn.relu(tf.matmul(pool2_vector, fc1_weights) + fc1_biases)
# 为了减少过拟合，加入Dropout层
keep_prob = tf.placeholder(tf.float32)
fc1_dropout = tf.nn.dropout(fc1, keep_prob)
# 第六层：全连接层
fc2_weights = tf.get_variable("fc2_weights", [1024, class_num],
initializer=tf.truncated_normal_initializer(
stddev=0.1))
# 神经元节点数1024, 分类节点10
fc2_biases = tf.get_variable(
"fc2_biases", [class_num], initializer=tf.constant_initializer(0.1))
fc2 = tf.matmul(fc1_dropout, fc2_weights) + fc2_biases
# 第七层：输出层
# softmax
y_conv = tf.nn.softmax(fc2)
# 定义交叉熵损失函数
cross_entropy = tf.reduce_mean(-tf.reduce_sum(y_ * tf.log(y_conv),
reduction_indices=[1]))
# 选择优化器，并让优化器最小化损失函数/收敛, 反向传播
train_step = tf.train.AdamOptimizer(1e-5).minimize(cross_entropy)
# tf.argmax()返回的是某一维度上其数据最大所在的索引值，在这里即代表预测值和真实值
# 判断预测值y和真实值y_中最大数的索引是否一致，y的值为1-class_num概率
correct_prediction = tf.equal(tf.argmax(y_conv, 1), tf.argmax(y_, 1))
# 用平均值来统计测试准确率
accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))
# 开始训练
saver = tf.train.Saver()
sess.run(tf.global_variables_initializer())
acc_train_train = []
acc_train_test = []
batch_size = 64
epoch_train = 50001 # restricted by the hardware in my computer
print("Training steps=" + str(epoch_train))
for i in range(epoch_train):
if (i*batch_size % x_train.shape[0]) > ((i + 1)*batch_size %
x_train.shape[0]):
x_data_train = np.vstack(
(x_train[i*batch_size % x_train.shape[0]:],
x_train[:(i+1)*batch_size % x_train.shape[0]]))
y_data_train = np.vstack(
(y_train[i*batch_size % y_train.shape[0]:],
y_train[:(i+1)*batch_size % y_train.shape[0]]))
x_data_test = np.vstack(
(x_test[i*batch_size % x_test.shape[0]:],
x_test[:(i+1)*batch_size % x_test.shape[0]]))
y_data_test = np.vstack(
(y_test[i*batch_size % y_test.shape[0]:],
y_test[:(i+1)*batch_size % y_test.shape[0]]))
else:
x_data_train = x_train[
i*batch_size % x_train.shape[0]:
(i+1)*batch_size % x_train.shape[0]]
y_data_train = y_train[
i*batch_size % y_train.shape[0]:
(i+1)*batch_size % y_train.shape[0]]
x_data_test = x_test[
i*batch_size % x_test.shape[0]:
(i+1)*batch_size % x_test.shape[0]]
y_data_test = y_test[
i*batch_size % y_test.shape[0]:
(i+1)*batch_size % y_test.shape[0]]
if i % 640 == 0:
train_accuracy = accuracy.eval(
feed_dict={x: x_data_train, y_: y_data_train, keep_prob: 1.0})
test_accuracy = accuracy.eval(
feed_dict={x: x_data_test, y_: y_data_test, keep_prob: 1.0})
print("step {}, training accuracy={}, testing accuracy={}".format(
i, train_accuracy, test_accuracy))
acc_train_train.append(train_accuracy)
acc_train_test.append(test_accuracy)
train_step.run(feed_dict={
x: x_data_train, y_: y_data_train, keep_prob: 0.5})
print("saving model...")
save_path = saver.save(sess, "./my_model/model.ckpt")
print("save model:{0} Finished".format(save_path))
batch_size_test = 64
epoch_test = y_test.shape[0] // batch_size_test + 1
acc_test = 0
for i in range(epoch_test):
if (i*batch_size_test % x_test.shape[0]) > ((i + 1)*batch_size_test %
x_test.shape[0]):
x_data_test = np.vstack((
x_test[i*batch_size_test % x_train.shape[0]:],
x_test[:(i+1)*batch_size_test % x_test.shape[0]]))
y_data_test = np.vstack((
y_test[i*batch_size_test % y_test.shape[0]:],
y_test[:(i+1)*batch_size_test % y_test.shape[0]]))
else:
x_data_test = x_test[
i*batch_size_test % x_test.shape[0]:
(i+1)*batch_size_test % x_test.shape[0]]
y_data_test = y_test[
i*batch_size_test % y_test.shape[0]:
(i+1)*batch_size_test % y_test.shape[0]]
# plt.imshow(x_data_test[0].reshape(28, 28), cmap="gray")
# plt.show()
# Calculate batch loss and accuracy
c = accuracy.eval(feed_dict={
x: x_data_test, y_: y_data_test, keep_prob: 1.0})
acc_test += c / epoch_test
print("{}-th test accuracy={}".format(i, acc_test))
print("At last, test accuracy={}".format(acc_test))
print("Finish!")
return acc_train_train, acc_train_test, acc_test
def plot_acc(acc_train_train, acc_train_test, acc_test):
plt.figure(1)
p1, p2 = plt.plot(list(range(len(acc_train_train))),
acc_train_train, 'r>',
list(range(len(acc_train_test))),
acc_train_test, 'b-')
plt.legend(handles=[p1, p2], labels=["training_acc", "testing_acc"])
plt.title("Accuracies During Training")
plt.show()
def main():
# integers: 4679
# alphabets: 9796
# Chinese_letters: 3974
# training_set : testing_set == 4 : 1
train_lst = ['alphabets', 'integers', 'alphabets',
'Chinese_letters', 'integers']
acc_train_train, acc_train_test, acc_test = lenet(train_lst)
plot_acc(acc_train_train, acc_train_test, acc_test)
if __name__ == '__main__':
main()