import tensorflow as tf
import os
os.environ['TF_CPP_MIN_LOG_LEVEL'] = '2'


def tensorflow_test():

    # a = 10
    # b = 20
    # c = a+b
    # print(c)

    con1 = tf.constant(10, name='a')
    con2 = tf.constant(20, name='b')
    sum_t = tf.add(con1, con2, name='sum')

    # 打印默认图
    # print(tf.get_default_graph())
    # print(con1.graph)
    # print(con2.graph)
    # print(sum_t.graph)
    # add(a, b)

    # 创建一个自定义的图
    # new_g = tf.Graph()
    # with new_g.as_default():
    #     new_a = tf.constant(30)
    #     new_b = tf.constant(40)
    #     new_sum = tf.add(new_a, new_b)
    # print(new_a.graph)
    # print(new_b.graph)
    # print(new_sum.graph)

    # print(sum)
    #     with tf.Session(graph=new_g) as sess:

    # 定义个placeholder()方法，实现函数传入参数功能
    a = tf.placeholder(tf.float32)
    b = tf.placeholder(tf.float32)
    sum_ab = tf.add(a, b)

    print(con1)
    print(con2)
    print(sum_t)
    with tf.Session(config=tf.ConfigProto(allow_soft_placement=True, log_device_placement=True)) as sess:
        # print(sess.run(sum_t))
        # 新图的资源
        # print(sess.run(sum_t))

        print(sess.run([con1, con2, sum_t]))
        print(con1.eval())

        # 把图序列化成events文件
        # file_writer = tf.summary.FileWriter("./tmp/summary/", graph=sess.graph)
        # print(sess.graph)

        # 运行placeholder
        print(sess.run([a, b, sum_ab], feed_dict={a: 3.0, b: 4.0}))
    return None


def shape_test():

    a_p = tf.placeholder(dtype=tf.float32, shape=[None, None])
    b_p = tf.placeholder(dtype=tf.float32, shape=[None, 10])
    c_p = tf.placeholder(dtype=tf.float32, shape=[3, 2])
    # 获取静态形状
    print("a_p的静态形状为：\n", a_p.get_shape())
    print("b_p的静态形状为：\n", b_p.get_shape())
    print("c_p的静态形状为：\n", c_p.get_shape())

    # 修改这些tensor形状
    # 静态修改
    a_p.set_shape([10, 12])
    print("a_p的静态形状为：\n", a_p.get_shape())
    b_p.set_shape([3, 10])
    print("b_p的静态形状为：\n", b_p.get_shape())
    # c_p.set_shape([2, 3])
    # print("c_p的静态形状为：\n", c_p.get_shape())
    # 形状固定，不能再用set_shape去修改，若能，set_shape修改张量本身的形状

    # 动态修改, 创建新的tensor，不会对原来的tensor修改
    c_reshape = tf.reshape(c_p, [2, 3])
    print("c_p的静态形状为：\n", c_p.get_shape())
    print("cc_reshape的静态形状为：\n", c_reshape.get_shape())

    # c_reshape = tf.reshape(c_p, [10, 3])


def variable_test():

    a = tf.Variable(initial_value=30.0)
    b = tf.Variable(initial_value=40.0)

    sum_t = tf.add(a, b)

    # 初始化变量
    init = tf.global_variables_initializer()

    print(a, b, sum_t)
    with tf.Session() as sess:
        sess.run(init)
        print(sess.run(sum_t))


tf.app.flags.DEFINE_integer('max_step', 0, "线性回归模型训练步数")

FLAGS = tf.app.flags.FLAGS


def linear_regression():
    """
    根据数据自实现一个线性回归过程
    :return:
    """
    # 1 准备好数据集：y = 0.8x + 0.7 100个样本
    # 100样本， [100 ,1] * [1, 1]
    with tf.variable_scope("original_data"):
        X = tf.random_normal([100, 1], mean=2, stddev=1, name='feature')

        y_true = tf.matmul(X, [[0.8]]) + 0.7

    # 2 建立线性模型
    # x [100 ,1] * [1, 1], y  [100]
    # y = wx + b , w[1, 1], b [1, ,1]
    # 随机初始化W1和b1
    # y = W·X + b，目标：求出权重W和偏置b
    # 训练模型参数必须使用tf.Variable
    with tf.variable_scope("linear_model"):
        weights = tf.Variable(initial_value=tf.random_normal([1, 1]), name='w')
        bias = tf.Variable(initial_value=tf.random_normal([1, 1]), name='b')
        y_predict = tf.matmul(X, weights) + bias

    with tf.variable_scope("loss"):
        # 3 确定损失函数（预测值与真实值之间的误差）-均方误差
        loss = tf.reduce_mean(tf.square(y_true - y_predict))

    # 4 梯度下降优化损失：需要指定学习率（超参数）
    with tf.variable_scope("optimizer"):
        optimizer = tf.train.GradientDescentOptimizer(learning_rate=0.01).minimize(loss)

    # 手动初始化
    init = tf.global_variables_initializer()

    # 收集观察的tensor效果
    tf.summary.scalar('losses', loss)
    tf.summary.histogram('w', weights)
    tf.summary.histogram('b', bias)

    # 合并Tensor
    merge = tf.summary.merge_all()

    # 添加一个saver
    saver = tf.train.Saver()

    # 开启会话运行整个模型
    with tf.Session() as sess:
        sess.run(init)

        # 训练线性回归模型
        # 优化之前打印初始化的模型参数
        print("随机初始化的权重为%f， 偏置为%f" % (weights.eval(), bias.eval()))
        # 第499步的误差为0.000243，权重为0.813589， 偏置为0.667376

        # 写入到events文件
        file_writer = tf.summary.FileWriter("./tmp/summary/linear/", graph=sess.graph)

        # 加载历史模型，基于历史模型训练
        saver.restore(sess, "./tmp/ckpt/linear/linear_regression.ckpt")

        for i in range(FLAGS.max_step):
            sess.run(optimizer)

            print("第%d步的误差为%f，权重为%f， 偏置为%f" % (i, loss.eval(), weights.eval(), bias.eval()))

            # 运行观察Tensor
            summary = sess.run(merge)
            file_writer.add_summary(summary, i)

            # 保存模型
            saver.save(sess, "./tmp/ckpt/linear/linear_regression.ckpt")


if __name__ == '__main__':
    # tensorflow_test()
    # shape_test()
    # variable_test()
    linear_regression()