import tensorflow as tf
tf.reset_default_graph() #流式计算图形graph  循环神经网络 将名字相同重置了图
import datetime #打印时间
import os   #保存文件
from tensorflow.examples.tutorials.mnist import input_data

# minst测试集
mnist = input_data.read_data_sets('../', one_hot=True)

# 每次使用100条数据进行训练
batch_size = 100
# 图像向量
width = 28
height = 28
# LSTM隐藏神经元数量
rnn_size = 256
# 输出层one-hot向量长度的
out_size = 10

def weight_variable(shape, w_alpha=0.01):
    initial = w_alpha * tf.random_normal(shape)
    return tf.Variable(initial)

def bias_variable(shape, b_alpha=0.1):
    initial = b_alpha * tf.random_normal(shape)
    return tf.Variable(initial)

# 权重及偏置
w = weight_variable([rnn_size, out_size])
b = bias_variable([out_size])

# 按照图片大小申请占位符
X = tf.placeholder(tf.float32, [None, height, width])
# 原排列[0,1,2]transpose为[1,0,2]代表前两维装置，如shape=(1,2,3)转为shape=(2,1,3)
# 这里的实际意义是把所有图像向量的相同行号向量转到一起，如x1的第一行与x2的第一行
x = tf.transpose(X, [1, 0, 2])
# reshape -1 代表自适应，这里按照图像每一列的长度为reshape后的列长度
x = tf.reshape(x, [-1, width])
# split默任在第一维即0 dimension进行分割，分割成height份，这里实际指把所有图片向量按对应行号进行重组
x = tf.split(x, height)

# LSTM
lstm_cell = tf.nn.rnn_cell.BasicLSTMCell(rnn_size)
# 这里RNN会有与输入层相同数量的输出层，我们只需要最后一个输出
outputs, status = tf.nn.static_rnn(lstm_cell, x, dtype=tf.float32)

#取最后一个进行矩阵乘法
y_conv = tf.add(tf.matmul(outputs[-1], w), b)
# 最小化损失优化
Y = tf.placeholder(dtype=tf.float32,shape = [None,10])
#损失使用的交叉熵
loss = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits_v2(logits=y_conv, labels=Y))
optimizer = tf.train.AdamOptimizer(0.01).minimize(loss)
# 计算准确率
correct = tf.equal(tf.argmax(y_conv, 1), tf.argmax(Y, 1))
accuracy = tf.reduce_mean(tf.cast(correct, tf.float32))

# 启动会话.开始训练
saver = tf.train.Saver()
session = tf.Session()
session.run(tf.global_variables_initializer())
step = 0
acc_rate = 0.90
while 1:
    batch_x, batch_y = mnist.train.next_batch(batch_size)
    batch_x = batch_x.reshape((batch_size, height, width))
    session.run(optimizer, feed_dict={X:batch_x,Y:batch_y})
    # 每训练10次测试一次
    if step % 10 == 0:
        batch_x_test = mnist.test.images
        batch_y_test = mnist.test.labels
        batch_x_test = batch_x_test.reshape([-1, height, width])
        acc = session.run(accuracy, feed_dict={X: batch_x_test, Y: batch_y_test})
        print(datetime.datetime.now().strftime('%c'), ' step:', step, ' accuracy:', acc)
        # 偏差满足要求，保存模型
        if acc >= acc_rate:
#             os.sep = ‘/’
            model_path = os.getcwd() + os.sep + str(acc_rate) + "mnist.model"
            saver.save(session, model_path, global_step=step)
            break
    step += 1
session.close()

import tensorflow as tf
tf.reset_default_graph()
from tensorflow.examples.tutorials.mnist import input_data

# 载入数据
mnist = input_data.read_data_sets("../", one_hot=True)

# 输入图片是28
n_input = 28
max_time = 28
lstm_size = 100  # 隐藏单元 可调
n_class = 10  # 10个分类
batch_size = 100   # 每次50个样本 可调
n_batch_size = mnist.train.num_examples // batch_size    # 计算一共有多少批次

# 这里None表示第一个维度可以是任意长度
# 创建占位符
x = tf.placeholder(tf.float32,[None, 28*28])
# 正确的标签
y = tf.placeholder(tf.float32,[None, 10])

# 初始化权重 ,stddev为标准差
weight = tf.Variable(tf.truncated_normal([lstm_size, n_class], stddev=0.1))
# 初始化偏置层
biases = tf.Variable(tf.constant(0.1, shape=[n_class]))

# 定义RNN网络
def RNN(X, weights, biases):
    #  原始数据为[batch_size,28*28]
    # input = [batch_size, max_time, n_input]
    input_ = tf.reshape(X,[-1, max_time, n_input])
    # 定义LSTM的基本单元
#     lstm_cell = tf.contrib.rnn.BasicLSTMCell(lstm_size)
    lstm_cell = tf.nn.rnn_cell.BasicLSTMCell(lstm_size)
    # final_state[0] 是cell state
    # final_state[1] 是hidden stat
    outputs, final_state = tf.nn.dynamic_rnn(lstm_cell, input_, dtype=tf.float32)
    display(final_state)
    results = tf.nn.softmax(tf.matmul(final_state[1],weights)+biases)
    return results
# 计算RNN的返回结果
prediction = RNN(x, weight, biases)
# 损失函数
loss = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(logits=prediction, labels=y))
# 使用AdamOptimizer进行优化
train_step = tf.train.AdamOptimizer(1e-4).minimize(loss)
# 将结果存下来
correct_prediction = tf.equal(tf.argmax(y, 1), tf.argmax(prediction, 1))
# 计算正确率
accuracy = tf.reduce_mean(tf.cast(correct_prediction,tf.float32))

saver = tf.train.Saver()

with tf.Session() as sess:
    sess.run(tf.global_variables_initializer())
    for epoch in range(6):
        for batch in range(n_batch_size):
            # 取出下一批次数据
            batch_xs,batch_ys = mnist.train.next_batch(batch_size)
            sess.run(train_step, feed_dict={x: batch_xs,y: batch_ys})
            if(batch%100==0):
                print(str(batch)+"/" + str(n_batch_size))
        acc = sess.run(accuracy, feed_dict={x: mnist.test.images, y: mnist.test.labels})
        print("Iter" + str(epoch) + " ,Testing Accuracy = " + str(acc))
        if acc >0.9:
            saver.save(sess,'./rnn_dynamic')
            break