TensorFlow 0.12.0 全连接层

TensorFlow 0.12.0 全连接层 (Batch * Width)

源代码

import tensorflow as tf
import numpy as np
import random

# 超参数
BATCH_SIZE = 5  # 批大小
MAX_EPOCH = 101  # 循环终止轮数


# 随机种子
def set_seed(manual_seed):
    random.seed(manual_seed)
    np.random.seed(manual_seed)
    tf.set_random_seed(manual_seed)


set_seed(0)


# 定义全连接网络层
def fc_layer(inputs, in_size, out_size, activation_function=None):
    """
    :param inputs: 1 x m
    :param in_size: m
    :param out_size: n
    :param activation_function:
    :return: 1 x n
    """
    Weights = tf.Variable(tf.random_normal([in_size, out_size]))
    biases = tf.Variable(tf.zeros([1, out_size]) + 0.1)
    Wx_plus_b = tf.matmul(inputs, Weights) + biases  # 加权求和层

    if activation_function is None:
        outputs = Wx_plus_b
    else:
        outputs = activation_function(Wx_plus_b)     # **函数层
    return outputs


if __name__ == '__main__':
    # 数据集读取(创建)
    x_data = np.linspace(-1, 1, BATCH_SIZE)[:, np.newaxis]      # data  shape=(B, W), B=BATCH_SIZE, W=1
    # print('x_data is:\n', x_data)
    # print('x_data shape:', x_data.shape)
    noise = np.random.normal(0, 0.05, x_data.shape)
    # print('n_data is:\n', noise)
    # print('n_data shape:', noise.shape)
    y_data = np.square(x_data) - 0.5 + noise                    # label
    # print('y_data is:\n', y_data)
    # print('y_data shape:', y_data.shape)

    # 模型架构，前向传播计算图
    input_x = tf.placeholder(dtype=tf.float32, shape=[None, 1])  # Batch * Width
    hidden_v = fc_layer(inputs=input_x, in_size=1, out_size=10, activation_function=tf.nn.relu)
    output_y = fc_layer(hidden_v, in_size=10, out_size=1, activation_function=None)

    # 权值初始化策略
    init = tf.global_variables_initializer()

    # 优化器，学习率设定
    optimizer = tf.train.GradientDescentOptimizer(learning_rate=0.1)

    # 目标函数，代价，损失，误差
    label_y = tf.placeholder(dtype=tf.float32, shape=[None, 1])  # Batch * Width
    # loss = tf.reduce_mean(tf.reduce_sum(tf.square(label_y - output_y), reduction_indices=[1]))
    loss = tf.reduce_mean(tf.square(label_y - output_y), reduction_indices=[0, 1])

    # 反向传播，梯度下降，权值更新策略
    update = optimizer.minimize(loss)

    # 建立会话
    sess = tf.Session()
    # 权值初始化
    sess.run(init)
    # 迭代训练
    for i in range(MAX_EPOCH):
        sess.run(update, feed_dict={input_x: x_data, label_y: y_data})
        # 打印过程信息
        if i % (MAX_EPOCH // 10) == 0:
            print('epoch =', i, ', loss =', sess.run(loss, feed_dict={input_x: x_data, label_y: y_data}))
        if i == MAX_EPOCH - 1:
            print('#######################################################################')
            print('########## epoch == {}: ##########'.format(i))
            print('output:', sess.run(output_y, feed_dict={input_x: x_data, label_y: y_data}))
            print('#######################################################################')
            print('error:', sess.run(label_y - output_y, feed_dict={input_x: x_data, label_y: y_data}))
            print('#######################################################################')
            print('square:', sess.run(tf.square(label_y - output_y), feed_dict={input_x: x_data, label_y: y_data}))
            print('#######################################################################')
            print('reduce_sum_axis1:', sess.run(tf.reduce_sum(tf.square(label_y - output_y), reduction_indices=[1]),
                                                feed_dict={input_x: x_data, label_y: y_data}))
            print('#######################################################################')
            print('reduce_mean:', sess.run(tf.reduce_mean(tf.reduce_sum(tf.square(label_y - output_y), reduction_indices=[1])),
                                           feed_dict={input_x: x_data, label_y: y_data}))
            print('#######################################################################')

运行结果

D:\Anaconda3\envs\py35tf012\python.exe E:/PySpace/TF0Study/全连接层BxW.py
epoch = 0 , loss = 0.661514
epoch = 10 , loss = 0.0903889
epoch = 20 , loss = 0.0460918
epoch = 30 , loss = 0.0280291
epoch = 40 , loss = 0.0194787
epoch = 50 , loss = 0.0149407
epoch = 60 , loss = 0.0121535
epoch = 70 , loss = 0.0101927
epoch = 80 , loss = 0.00867454
epoch = 90 , loss = 0.00743103
epoch = 100 , loss = 0.0063873
#######################################################################
########## epoch == 100: ##########
output: [[ 0.50809157]
 [-0.09274025]
 [-0.51904196]
 [-0.09954993]
 [ 0.5692001 ]]
#######################################################################
error: [[ 0.08011103]
 [-0.13725188]
 [ 0.06797886]
 [-0.0384054 ]
 [ 0.02417779]]
#######################################################################
square: [[ 0.00641778]
 [ 0.01883808]
 [ 0.00462113]
 [ 0.00147498]
 [ 0.00058457]]
#######################################################################
reduce_sum_axis1: [ 0.00641778  0.01883808  0.00462113  0.00147498  0.00058457]
#######################################################################
reduce_mean: 0.0063873
#######################################################################

Process finished with exit code 0

运行环境：
Python=3.5
TensorFlow=0.12.0

深度学习训练算法五大步骤

1、数据
2、模型
3、优化器
4、目标函数
5、迭代训练

基于静态图的TensorFlow的会话机制（ tf.Session() ）

“会话是TensorFlow的主要交互方式。一般而言，TensorFlow处理数据的流程是：建立会话、生成一张空图、添加各个结点和边，形成一个有连接点的图，然后启动图，进行系统的执行。”
——摘自：
TensorFlow 深度学习应用实践 / 王晓华著. — 北京：清华大学出版社，2018
ISBN 978-7-302-48795-1
(90页)

建立并使用tf.Session的方法：

# tf 0.12.0

import tensorflow as tf

matrix1 = tf.constant([[3, 3]])
matrix2 = tf.constant([[2],
                       [2]])

product = tf.matmul(matrix1, matrix2)

# ####################### method 1
flag = 0
# flag = 1
if flag:
    sess = tf.Session()
    result = sess.run(product)
    print('method 1:', result)
    sess.close()

# ####################### method 2
# flag = 0
flag = 1
if flag:
    with tf.Session() as sess:
        result = sess.run(product)
        print('method 2:', result)

Many thanks to Teacher MorvanZhou and TingsongYu!