tensorflow_用多层感知器模型训练MNIST数据集
import warnings import tensorflow as tf from tensorflow.examples.tutorials.mnist import input_data warnings.filterwarnings('ignore') # 导入数据mnist path = './mnist/input_data' mnist = input_data.read_data_sets(path, one_hot=True) # 定义网络参数 learning_rate = 0.01 iterations = 360 batch_number = 100 display_step = 10 # 定义占位符 # 因为这里的图片是28*28，所以是784个特征字段，输出为0-->9，所以Y输出为10列 X = tf.placeholder(tf.float32, shape=[None, 784], name='X') Y = tf.placeholder(tf.float32, shape=[None, 10], name='Y') # 定义隐层的对应的神经元数量 hidden_layer1 = 4096 hidden_layer2 = 2048 # 变量初始化 t = tf.truncated_normal_initializer(mean=0.21, stddev=0.1, seed=7) weight = { # 'w_layer1': tf.get_variable('w_layer1', [784, hidden_layer1], initializer=t), "w_layer1": tf.Variable(tf.random_normal([784, hidden_layer1])), # 'out_layer': tf.get_variable('out_layer', [hidden_layer1, 10], initializer=t), "out_layer": tf.Variable(tf.random_normal([hidden_layer1, 10])) } bias = { # 'bias1': tf.Variable(tf.truncated_normal([hidden_layer1], mean=0.21, stddev=0.1, seed=7)), # 'bias_out': tf.Variable(tf.truncated_normal([10], mean=0.21, stddev=0.1, seed=7)), 'bias1': tf.Variable(tf.random_normal([hidden_layer1])), 'bias_out': tf.Variable(tf.random_normal([10])), } weight['w_layer1'].shape out[7]: TensorShape([Dimension(784), Dimension(4096)]) init = tf.global_variables_initializer() with tf.Session() as sess: sess.run(init) print(weight['w_layer1'][0].eval()) print(bias['bias_out'].eval()) sess.close() out[8]: [-0.9179551 2.620515 0.84942716 ... 1.2998239 -0.58302623 -0.2974956 ] [ 1.1626408 -0.7264125 0.1215397 0.47657487 -0.20571353 0.42560428 1.0548028 -2.1683683 -0.8865439 -0.28622103] # 构建多层感知器模型 def multilayer_perceptron(x, w, b): hidden_layer_one = tf.matmul(x , w['w_layer1']) + b['bias1'] hidden_layer_one = tf.nn.relu(hidden_layer_one) out_layer = tf.matmul(hidden_layer_one, w['out_layer']) + b['bias_out'] return out_layer # 调用多层感知器模型，获得预测值 y_pred = multilayer_perceptron(X, weight, bias) # 定义损失代价函数 tf.reduce_mean()计算平均值 cross_entropys = tf.nn.softmax_cross_entropy_with_logits_v2(labels=Y, logits=y_pred) cross_entropy = tf.reduce_mean(cross_entropys) # 定义优化函数，进行"梯度下降" optimizer = tf.train.AdamOptimizer(learning_rate).minimize(cross_entropy) init = tf.global_variables_initializer() with tf.Session() as sess: sess.run(init) for iteration in range(iterations): # 初始训练误差，计算每轮批量迭代次数 train_cost=0 batch_times=int(mnist.train._num_examples/batch_number) for i in range(batch_times): # 每次取batch_number张图 batch_X, batch_Y = mnist.train.next_batch(batch_number) # 运行优化函数 # 这里返回一个[optimizer,cost]的list, 其中 _代表optimizer,batch_cost代表cost的值 _,batch_cost = sess.run([optimizer,cross_entropy],feed_dict={X:batch_X, Y:batch_Y}) # 返回训练集误差：每次计算batch_number张图的batch_cost，计算了i次，所以最后除以batch_numbers train_cost += batch_cost / batch_times if iteration % display_step==0: # %04d: % 转义说明符 ; 0 指以0填充前面的位数 ；4 四位数； d 十进制整数 # "{:.9f}".format(train_cost) 以保留小数点后9位显示train_cost prediction=tf.equal(tf.argmax(y_pred, 1), tf.argmax(Y,1)) accuracy=tf.reduce_mean(tf.cast(prediction,"float")) print("Epoch:","%04d"%(iteration+1), "Train_cost :","{:.9f} ".format(train_cost), "accuracy :", sess.run(accuracy,feed_dict={X:mnist.test.images, Y:mnist.test.labels})) # print("accuracy :",sess.run(accuracy,feed_dict={X:mnist.test.images, Y:mnist.test.labels})) # tf.arg_max(pred,1):得到向量中最大数的下标，1代表水平方向 # tf.equal():返回布尔值，相等返回1，否则0 # 最后返回大小[none,1]的向量，1所在位置为布尔类型数据 prediction=tf.equal(tf.argmax(y_pred, 1), tf.argmax(Y,1)) # tf.cast():将布尔型向量转换成浮点型向量 # tf.reduce_mean():求所有数的均值 # 返回正确率：也就是所有为1的数目占所有数目的比例 accuracy=tf.reduce_mean(tf.cast(prediction,"float")) # 打印正确率 print("Train :",sess.run(accuracy,feed_dict={X:mnist.train.images,Y:mnist.train.labels})) print("Test :",sess.run(accuracy,feed_dict={X:mnist.test.images,Y:mnist.test.labels})) sess.close()
为了看清训练过程的一些细节，可以观看如下代码及其结果：
with tf.Session() as sess: sess.run(tf.global_variables_initializer()) batch_X, batch_Y = mnist.train.next_batch(100) test_y = sess.run(y_pred, feed_dict={X:batch_X}) print(type(test_y)) print(test_y[0]) print(tf.argmax(test_y, 1)) print(tf.argmax(test_y, 1).eval()) print(tf.argmax(batch_Y, 1).eval()) print("原始的标签数据", batch_Y) sess.close()
#运行结果： <class 'numpy.ndarray'> [ 766.4617 -133.16173 604.783 758.3629 595.90173 -153.06392 321.0015 -835.967 -668.4424 -585.8179 ] Tensor("ArgMax_74:0", shape=(100,), dtype=int64) [0 4 4 6 4 4 0 2 4 0 4 4 4 0 0 3 4 0 4 4 0 3 4 0 4 0 0 0 4 2 4 4 4 4 4 4 4 0 2 4 0 0 4 4 0 4 4 0 4 3 4 3 0 2 4 6 4 0 0 0 5 0 4 0 2 0 0 0 0 0 4 0 2 3 3 4 0 6 4 4 0 4 4 2 3 4 0 4 0 0 4 4 0 0 4 0 4 2 0 4] [3 8 8 6 4 8 4 5 7 8 0 3 4 3 8 5 8 3 0 9 6 7 5 7 9 8 3 4 0 3 1 8 5 5 3 4 0 1 8 8 7 9 3 7 8 8 3 2 7 7 2 6 4 4 3 7 8 8 8 0 8 5 0 6 8 6 8 5 7 4 1 4 4 4 1 9 9 7 5 8 2 3 7 6 6 7 7 4 4 6 7 1 0 1 5 9 8 4 2 0] 原始的标签数据 [[0. 0. 0. 1. 0. 0. 0. 0. 0. 0.] [0. 0. 0. 0. 0. 0. 0. 0. 1. 0.] [0. 0. 0. 0. 0. 0. 0. 0. 1. 0.] [0. 0. 0. 0. 0. 0. 1. 0. 0. 0.] [0. 0. 0. 0. 1. 0. 0. 0. 0. 0.] [0. 0. 0. 0. 0. 0. 0. 0. 1. 0.] [0. 0. 0. 0. 1. 0. 0. 0. 0. 0.] [0. 0. 0. 0. 0. 1. 0. 0. 0. 0.] [0. 0. 0. 0. 0. 0. 0. 1. 0. 0.] [0. 0. 0. 0. 0. 0. 0. 0. 1. 0.] [1. 0. 0. 0. 0. 0. 0. 0. 0. 0.] [0. 0. 0. 1. 0. 0. 0. 0. 0. 0.] [0. 0. 0. 0. 1. 0. 0. 0. 0. 0.] [0. 0. 0. 1. 0. 0. 0. 0. 0. 0.] [0. 0. 0. 0. 0. 0. 0. 0. 1. 0.] [0. 0. 0. 0. 0. 1. 0. 0. 0. 0.] [0. 0. 0. 0. 0. 0. 0. 0. 1. 0.] [0. 0. 0. 1. 0. 0. 0. 0. 0. 0.] [1. 0. 0. 0. 0. 0. 0. 0. 0. 0.] [0. 0. 0. 0. 0. 0. 0. 0. 0. 1.] [0. 0. 0. 0. 0. 0. 1. 0. 0. 0.] [0. 0. 0. 0. 0. 0. 0. 1. 0. 0.] [0. 0. 0. 0. 0. 1. 0. 0. 0. 0.] [0. 0. 0. 0. 0. 0. 0. 1. 0. 0.] [0. 0. 0. 0. 0. 0. 0. 0. 0. 1.] [0. 0. 0. 0. 0. 0. 0. 0. 1. 0.] [0. 0. 0. 1. 0. 0. 0. 0. 0. 0.] [0. 0. 0. 0. 1. 0. 0. 0. 0. 0.] [1. 0. 0. 0. 0. 0. 0. 0. 0. 0.] [0. 0. 0. 1. 0. 0. 0. 0. 0. 0.] [0. 1. 0. 0. 0. 0. 0. 0. 0. 0.] [0. 0. 0. 0. 0. 0. 0. 0. 1. 0.] [0. 0. 0. 0. 0. 1. 0. 0. 0. 0.] [0. 0. 0. 0. 0. 1. 0. 0. 0. 0.] [0. 0. 0. 1. 0. 0. 0. 0. 0. 0.] [0. 0. 0. 0. 1. 0. 0. 0. 0. 0.] [1. 0. 0. 0. 0. 0. 0. 0. 0. 0.] [0. 1. 0. 0. 0. 0. 0. 0. 0. 0.] [0. 0. 0. 0. 0. 0. 0. 0. 1. 0.] [0. 0. 0. 0. 0. 0. 0. 0. 1. 0.] [0. 0. 0. 0. 0. 0. 0. 1. 0. 0.] [0. 0. 0. 0. 0. 0. 0. 0. 0. 1.] [0. 0. 0. 1. 0. 0. 0. 0. 0. 0.] [0. 0. 0. 0. 0. 0. 0. 1. 0. 0.] [0. 0. 0. 0. 0. 0. 0. 0. 1. 0.] [0. 0. 0. 0. 0. 0. 0. 0. 1. 0.] [0. 0. 0. 1. 0. 0. 0. 0. 0. 0.] [0. 0. 1. 0. 0. 0. 0. 0. 0. 0.] [0. 0. 0. 0. 0. 0. 0. 1. 0. 0.] [0. 0. 0. 0. 0. 0. 0. 1. 0. 0.] [0. 0. 1. 0. 0. 0. 0. 0. 0. 0.] [0. 0. 0. 0. 0. 0. 1. 0. 0. 0.] [0. 0. 0. 0. 1. 0. 0. 0. 0. 0.] [0. 0. 0. 0. 1. 0. 0. 0. 0. 0.] [0. 0. 0. 1. 0. 0. 0. 0. 0. 0.] [0. 0. 0. 0. 0. 0. 0. 1. 0. 0.] [0. 0. 0. 0. 0. 0. 0. 0. 1. 0.] [0. 0. 0. 0. 0. 0. 0. 0. 1. 0.] [0. 0. 0. 0. 0. 0. 0. 0. 1. 0.] [1. 0. 0. 0. 0. 0. 0. 0. 0. 0.] [0. 0. 0. 0. 0. 0. 0. 0. 1. 0.] [0. 0. 0. 0. 0. 1. 0. 0. 0. 0.] [1. 0. 0. 0. 0. 0. 0. 0. 0. 0.] [0. 0. 0. 0. 0. 0. 1. 0. 0. 0.] [0. 0. 0. 0. 0. 0. 0. 0. 1. 0.] [0. 0. 0. 0. 0. 0. 1. 0. 0. 0.] [0. 0. 0. 0. 0. 0. 0. 0. 1. 0.] [0. 0. 0. 0. 0. 1. 0. 0. 0. 0.] [0. 0. 0. 0. 0. 0. 0. 1. 0. 0.] [0. 0. 0. 0. 1. 0. 0. 0. 0. 0.] [0. 1. 0. 0. 0. 0. 0. 0. 0. 0.] [0. 0. 0. 0. 1. 0. 0. 0. 0. 0.] [0. 0. 0. 0. 1. 0. 0. 0. 0. 0.] [0. 0. 0. 0. 1. 0. 0. 0. 0. 0.] [0. 1. 0. 0. 0. 0. 0. 0. 0. 0.] [0. 0. 0. 0. 0. 0. 0. 0. 0. 1.] [0. 0. 0. 0. 0. 0. 0. 0. 0. 1.] [0. 0. 0. 0. 0. 0. 0. 1. 0. 0.] [0. 0. 0. 0. 0. 1. 0. 0. 0. 0.] [0. 0. 0. 0. 0. 0. 0. 0. 1. 0.] [0. 0. 1. 0. 0. 0. 0. 0. 0. 0.] [0. 0. 0. 1. 0. 0. 0. 0. 0. 0.] [0. 0. 0. 0. 0. 0. 0. 1. 0. 0.] [0. 0. 0. 0. 0. 0. 1. 0. 0. 0.] [0. 0. 0. 0. 0. 0. 1. 0. 0. 0.] [0. 0. 0. 0. 0. 0. 0. 1. 0. 0.] [0. 0. 0. 0. 0. 0. 0. 1. 0. 0.] [0. 0. 0. 0. 1. 0. 0. 0. 0. 0.] [0. 0. 0. 0. 1. 0. 0. 0. 0. 0.] [0. 0. 0. 0. 0. 0. 1. 0. 0. 0.] [0. 0. 0. 0. 0. 0. 0. 1. 0. 0.] [0. 1. 0. 0. 0. 0. 0. 0. 0. 0.] [1. 0. 0. 0. 0. 0. 0. 0. 0. 0.] [0. 1. 0. 0. 0. 0. 0. 0. 0. 0.] [0. 0. 0. 0. 0. 1. 0. 0. 0. 0.] [0. 0. 0. 0. 0. 0. 0. 0. 0. 1.] [0. 0. 0. 0. 0. 0. 0. 0. 1. 0.] [0. 0. 0. 0. 1. 0. 0. 0. 0. 0.] [0. 0. 1. 0. 0. 0. 0. 0. 0. 0.] [1. 0. 0. 0. 0. 0. 0. 0. 0. 0.]]
清澈的爱，只为中国
