Create a simple deep neural network just with one hidden layer. reached 95.1% accuracy

deep_neural_network_with_hidden_layer.py

from tensorflow.examples.tutorials.mnist import input_data

# 如果 one_hot 设为 False , 标签则会输出数字
mnist = input_data.read_data_sets("../MNIST_data", one_hot=True)
import tensorflow as tf

batch_size = 50
display_step = 2
learning_rate = 0.03
training_epochs = 30


x = tf.placeholder(tf.float32, shape=[None, 784])
y = tf.placeholder(tf.float32, shape=[None, 10])

# 创建输入层（input layer）和隐藏层（hidden layer）的weight和bias变量
w = tf.Variable(tf.truncated_normal([784, 512]), name='w')
b = tf.Variable(tf.zeros([512]), name='b')

# 创建隐藏层（hidden layer）和输出层（output layer）的weight和bias变量
w1 = tf.Variable(tf.truncated_normal([512, 10]), name='w1')
b1 = tf.Variable(tf.zeros([10]), name='b1')

# 计算隐藏层（hidden layer），添加 sigmoid 激活函数
h_z = tf.nn.bias_add(tf.matmul(x, w), b)
h_a = tf.sigmoid(h_z)

# 计算输出层（output layer）假设值，后面会调用 softmax_cross_entropy_with_logits
# 所以不用添加 softmax 激活函数
o_z = tf.nn.bias_add(tf.matmul(h_a,w1), b1)

cost = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(logits=o_z, labels=y))

optimizer = tf.train.AdamOptimizer(learning_rate).minimize(cost)

correct_prediction = tf.equal(tf.argmax(o_z,1), tf.argmax(y,1))
accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))

init = tf.global_variables_initializer()
with tf.Session() as sess:
    sess.run(init)
    for epoch in range(training_epochs):
        avg_cost = 0.
        total_batch = mnist.train.num_examples // batch_size
        for i in range(total_batch):
            batch_images, batch_labels = mnist.train.next_batch(batch_size)
            _, c = sess.run([optimizer, cost], feed_dict={x: batch_images,
                                                          y: batch_labels})
            avg_cost += c / total_batch
        if (epoch +1) % display_step == 0:
            acc = sess.run(accuracy, feed_dict={x: mnist.test.images[:3000],
                                                y: mnist.test.labels[:3000]})
            print("Epoch: ", '%04d' % (epoch+1), "cost=", "{:.9f}".format(avg_cost),
                "accuracy : {:.3f}".format(acc))

    print("Optimization Finished!")
    saver = tf.train.Saver()
    save_path = saver.save(sess, "./sess.p")