import tensorflow as tf from dataset import * import time #时间标记起始点 time.clock() #导入数据 x_train, y_train, x_validating, y_validating, x_test, y_test = data_set() #定义参数 layer_dimensions = [784, 30, 50, 25, 33, 25, 15, 20, 10] regularization_rate = 0.001 w_cache, b_cache = initialization(layer_dimensions) #数据输入接口 x = tf.placeholder(tf.float32, shape=(None,784), name='x_input') y_ = tf.placeholder(tf.float32, shape=(None,10), name='y_input') #前向传播 y, regularizer_cache = propagation_forward(x, w_cache, b_cache, regularization_rate) #普通损失函数,均方失真 cross_entropy_mean = tf.nn.sparse_softmax_cross_entropy_with_logits( logits=y, labels=tf.argmax(y_, 1)) #logits为未归一化的量 loss = cross_entropy_mean + tf.accumulate_n(regularizer_cache) #优化方法 global_step = tf.Variable(0) learning_rate = tf.train.exponential_decay(0.01, global_step, 100, 0.99, staircase=True) train_step = tf.train.AdamOptimizer(learning_rate).minimize(loss, global_step=global_step) #正确度计算 correct_prediction = tf.equal(tf.arg_max(y, 1), tf.arg_max(y_, 1)) accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32)) #创建会话 with tf.Session() as sess: #初始化变量 init_op = tf.global_variables_initializer() sess.run(init_op) STEPS = 50000 batch_size = 256 x_axis = [i for i in range(STEPS+1) if (i % 100 == 0)] train_line = [] validating_line = [] test_line = [] for i in range(STEPS+1): xs, ys = batch(x_train, y_train, batch_size, i) sess.run(train_step, feed_dict={x: xs, y_: ys}) #计算正确率 if i % 100 == 0: train_acc = sess.run(accuracy, feed_dict={x: x_train, y_: y_train}) validating_acc = sess.run(accuracy, feed_dict={x: x_validating, y_: y_validating}) test_acc = sess.run(accuracy, feed_dict={x: x_test, y_: y_test}) train_line.append(train_acc) validating_line.append(validating_acc) test_line.append(test_acc) if i % 1000 == 0: print("after %d train steps, train accuracy on model is:%s" % (i, train_acc)) #print("after %d train steps, validating accuracy on model is:%s" % (i, validating_acc)) print("after %d train steps, test accuracy on model is:%s" % (i, test_acc)) #输出运行时间 print("total_time is:%s" % time.clock()) #性能曲线图 draw(x_axis, train_line, validating_line, test_line)
辅助函数:
from tensorflow.examples.tutorials.mnist import input_data import tensorflow as tf import matplotlib.pyplot as plt import numpy as np def data_set(): mnist = input_data.read_data_sets("D:/DL_dataset/TensorFlow/", one_hot=True) x_train = mnist.train.images #(55000,784) 图片为28*28=784像素,以下也是 y_train = mnist.train.labels #(55000,10) x_validating = mnist.validation.images #(5000,784) y_validating = mnist.validation.labels #(5000,10) x_test = mnist.test.images #(10000,784) y_test = mnist.test.labels #(10000,10) return x_train, y_train, x_validating, y_validating, x_test, y_test def initialization(layer_dimensions): w_cache = [] b_cache = [] for i in range(1, len(layer_dimensions)): w = tf.Variable(tf.random_normal((layer_dimensions[i-1], layer_dimensions[i]), stddev=1, seed=1)) b = tf.Variable(tf.constant(0., shape=[1, layer_dimensions[i]])) w_cache.append(w) b_cache.append(b) return w_cache, b_cache #前向传播与正则项计算 def propagation_forward(x, w_cache, b_cache, regularization_rate): a = x long = len(w_cache) regularizer_cache = [] regularizer = tf.contrib.layers.l2_regularizer(regularization_rate) if long> 1: for i in range(long-1): a = tf.nn.tanh(tf.matmul(a, w_cache[i]) + b_cache[i]) regularizer_cache.append(regularizer(w_cache[i])) a = tf.matmul(a, w_cache[long-1]) + b_cache[long-1] regularizer_cache.append(regularizer(w_cache[long-1])) return a, regularizer_cache #画性能曲线 def draw(total_iteration, train, validating, test): plt.figure(figsize=(8, 6), dpi=80) plt.subplot(1, 1, 1) xmin, xmax = min(total_iteration), max(total_iteration) ymin, ymax = min(train), max(train) #dx = (xmax - xmin) * 0.2 #dy = (ymax - ymin) * 0.2 plt.plot(total_iteration, train, '-y', label='train_acc') plt.plot(total_iteration, validating, '-r', label='valicating_acc') plt.plot(total_iteration, test, '-g', label='test_acc') plt.legend(loc='lower right') # 设置轴记号 plt.xticks(np.linspace(0, xmax, 11, endpoint=True)) plt.yticks(np.linspace(0, 1.0, 11, endpoint=True)) plt.show() #设置批量 def batch(x_train, y_train, batch_size, i): #整除运算,返回批数量 batch_nums = x_train.shape[0] // batch_size i = i % batch_size #舍弃掉最后批整数倍以外的多余数据 start = batch_size * i end = batch_size * (i + 1) - 1 xs = x_train[start:end] ys = y_train[start:end] return xs, ys
结果:训练集与测试集正确率分别为:99.83%, 97.39%