目标
本文旨在介绍 tensorflow 入门知识点及实战示例，希望各位新手同学能在学习之后熟练 tensorflow 相关基本操作
简单的循环神经网络
import tensorflow as tf
from tensorflow.examples.tutorials.mnist import input_data
mnist = input_data.read_data_sets('MNIST', one_hot=True)
batch_size = 64
n_batches = mnist.train.num_examples // batch_size
n_classes = 10 # 类别个数
hidden_size = 128 # 隐层纬度
steps = 28 # 最大序列
embedding_size = 28 # 输入维度
x = tf.placeholder(tf.float32, [None, 784])
y = tf.placeholder(tf.float32, [None, 10])
weights = tf.Variable(tf.random_normal([hidden_size, n_classes], stddev=0.1))
biases = tf.Variable(tf.zeros([n_classes]))
def RNN(x, w, b):
inputs = tf.reshape(x, shape = [-1, steps, embedding_size]) # 将图像拉成一个时序序列
cell = tf.contrib.rnn.BasicRNNCell(hidden_size) # 每个 RNN 隐藏输出维度 hidden_size
_, final_state = tf.nn.dynamic_rnn(cell, inputs, dtype=tf.float32) # rnn 模型计算获取最有状态向量
result = tf.nn.softmax(tf.matmul(final_state, w) + b) # 对结果进行 softmax
return result
predict = RNN(x, weights, biases) # 预测结果
loss = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(labels=y, logits=predict)) # 损失
opt = tf.train.AdamOptimizer(0.001).minimize(loss) # 定义优化器
correct = tf.equal(tf.argmax(y,1), tf. argmax(predict,1))
accuracy = tf.reduce_mean(tf.cast(correct, tf.float32)) # 准确率
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
total_batch = 0
last_batch = 0
best = 0
for epoch in range(100):
for _ in range(n_batches):
xx, yy = mnist.train.next_batch(batch_size)
sess.run(opt, {x:xx, y:yy})
acc, l = sess.run([accuracy, loss], {x:mnist.test.images, y:mnist.test.labels})
if acc > best:
best = acc
last_batch = total_batch
print('eopch:%d, acc:%f, loss:%f'%(epoch, acc, l))
if total_batch - last_batch > 5: # 早停条件
print('early stop')
break
total_batch += 1
结果输出
eopch:0, acc:0.878200, loss:1.589309
eopch:1, acc:0.907200, loss:1.556449
eopch:2, acc:0.917600, loss:1.546643
eopch:4, acc:0.933500, loss:1.528619
eopch:5, acc:0.950100, loss:1.512312
eopch:6, acc:0.951100, loss:1.511004
early stop
要点一
关于循环神经网络及其变体的基础知识大家可以上网去找，很多学习资料，也可以参考我之前写的文章学习：juejin.cn/post/697234…
要点二
相较于上一个单纯的多隐层网络 + Dropout 只有 97.8% 的准确率，使用循环神经网络准确率只有 95% ，而且也没有卷积神经网络的准确率高，说明 CNN 在处理图像方面天生就有优势，循环神经网络适合处理文本数据。
