【TensorFlow】随机训练和批训练的比较与实现

一、随机训练和批训练

随机训练：一次随机抽样训练数据和目标数据对完成训练。
批训练：一次大批量训练取平均损失来进行梯度计算，批量训练大小可以一次上扩到整个数据集。
批训练和随机训练的差异：优化器方法和收敛的不同
批训练的难点在于：确定合适的batch_size
二者比较

训练类型	优点	缺点
随机训练	脱离局部最小	一般需要更多的迭代次数才收敛
批训练	快速得到最小损失	耗费更多的计算资源

二、实现随机训练

import numpy as np
import tensorflow as tf
import matplotlib.pyplot as plt
from tensorflow.python.framework import ops
ops.reset_default_graph()
# 一、随机训练:# 1.创建计算图
sess = tf.Session()# 2.创建数据
x_vals = np.random.normal(1, 0.1, 100)
y_vals = np.repeat(10., 100)
x_data = tf.placeholder(shape=[1], dtype=tf.float32)
y_target = tf.placeholder(shape=[1], dtype=tf.float32)# 3.创建变量
A = tf.Variable(tf.random_normal(shape=[1]))# 4.增加图操作
my_output = tf.multiply(x_data, A)# 5.声明L2正则损失
loss = tf.square(my_output - y_target)# 6.声明优化器 学习率为0.02
my_opt = tf.train.GradientDescentOptimizer(0.02)
train_step = my_opt.minimize(loss)# 7.初始化变量
init = tf.global_variables_initializer()
sess.run(init)# 8.保存loss数据用于绘图
loss_stochastic = []# 9.开始训练
for i in range(100):rand_index = np.random.choice(100)rand_x = [x_vals[rand_index]]rand_y = [y_vals[rand_index]]sess.run(train_step, feed_dict={x_data: rand_x, y_target: rand_y})if (i+1)%5==0:print('Step #' + str(i+1) + ' A = ' + str(sess.run(A)))temp_loss = sess.run(loss, feed_dict={x_data: rand_x, y_target: rand_y})print('Loss = ' + str(temp_loss))loss_stochastic.append(temp_loss)

# 输出结果
Step #5 A = [2.0631378]
Loss = [60.90259]
Step #10 A = [3.560384]
Loss = [35.39518]
Step #15 A = [4.7225595]
Loss = [37.812637]
Step #20 A = [5.681144]
Loss = [13.796157]
Step #25 A = [6.4919457]
Loss = [13.752169]
Step #30 A = [7.1609416]
Loss = [9.70855]
Step #35 A = [7.710085]
Loss = [5.826261]
Step #40 A = [8.253489]
Loss = [7.3934216]
Step #45 A = [8.671478]
Loss = [2.5475926]
Step #50 A = [8.993064]
Loss = [1.32571]
Step #55 A = [9.101872]
Loss = [0.67589337]
Step #60 A = [9.256593]
Loss = [5.34419]
Step #65 A = [9.329251]
Loss = [0.58555096]
Step #70 A = [9.421848]
Loss = [3.088755]
Step #75 A = [9.563117]
Loss = [6.0601945]
Step #80 A = [9.661991]
Loss = [0.05205128]
Step #85 A = [9.8208685]
Loss = [2.3963788]
Step #90 A = [9.8652935]
Loss = [0.19284673]
Step #95 A = [9.842097]
Loss = [4.9211507]
Step #100 A = [10.044914]
Loss = [4.2354054]

三、实现批训练

import numpy as np
import tensorflow as tf
import matplotlib as pltfrom tensorflow.python.framework import ops
ops.reset_default_graph()sess = tf.Session()# 1.声明批量大小（一次传入多少训练数据）
batch_size = 20# 2.声明模型的数据、占位符和变量。
# 这里能做的是改变占位符的形状，占位符有两个维度：
# 第一个维度为None，第二个维度是批量训练中的数据量。
# 我们能显式地设置维度为20，也能设为None。
# 我们必须知道训练模型中的维度，从而阻止不合法的矩阵操作
x_vals = np.random.normal(1,0.1,100)
y_vals = np.repeat(10.,100)
x_data = tf.placeholder(shape=[None, 1], dtype=tf.float32)
y_target = tf.placeholder(shape=[None, 1], dtype=tf.float32)
A = tf.Variable(tf.random_normal(shape=[1,1]))# 3.现在在计算图中增加矩阵乘法操作，
# 切记矩阵乘法不满足交换律，所以在matmul()函数中的矩阵参数顺序要正确：
my_output = tf.multiply(x_data, A)# 4.改变损失函数
# 批量训练时损失函数是每个数据点L2损失的平均值
loss = tf.reduce_mean(tf.square(my_output - y_target))# 5.声明优化器
my_opt = tf.train.GradientDescentOptimizer(0.02)
train_step = my_opt.minimize(loss)# 6.在训练中通过循环迭代优化模型算法。
# 为了绘制损失值图与随机训练对比
# 这里初始化一个列表每间隔5次迭代保存损失函数# 初始化变量
init = tf.global_variables_initializer()
sess.run(init)loss_batch = []
for i in range(100):# 每次用0~100中取20个数作为索引值rand_index = np.random.choice(100, size=batch_size)# 转置rand_x = np.transpose([x_vals[rand_index]])rand_y = np.transpose([y_vals[rand_index]])sess.run(train_step, feed_dict={x_data:rand_x,y_target:rand_y})if (i+1)%5 == 0:print("Step # " + str(i+1) + ' A = ' + str(sess.run(A)))temp_loss = sess.run(loss, feed_dict={x_data:rand_x,y_target:rand_y})print('Loss = ' + str(temp_loss))loss_batch.append(temp_loss)

在这里插入代码片

# 输出结果
Step # 5 A = [[2.626382]]
Loss = 55.444374
Step # 10 A = [[3.980196]]
Loss = 36.855064
Step # 15 A = [[5.0858808]]
Loss = 22.765038
Step # 20 A = [[5.9751787]]
Loss = 15.496961
Step # 25 A = [[6.713659]]
Loss = 12.349718
Step # 30 A = [[7.2950797]]
Loss = 7.5467796
Step # 35 A = [[7.782353]]
Loss = 5.17468
Step # 40 A = [[8.20625]]
Loss = 4.1199327
Step # 45 A = [[8.509094]]
Loss = 2.6329637
Step # 50 A = [[8.760488]]
Loss = 1.9998455
Step # 55 A = [[8.967735]]
Loss = 1.6577679
Step # 60 A = [[9.1537]]
Loss = 1.4356906
Step # 65 A = [[9.317189]]
Loss = 1.9666836
Step # 70 A = [[9.387019]]
Loss = 1.9287064
Step # 75 A = [[9.499526]]
Loss = 1.7477573
Step # 80 A = [[9.594302]]
Loss = 1.719229
Step # 85 A = [[9.666611]]
Loss = 1.4769726
Step # 90 A = [[9.711805]]
Loss = 1.1235845
Step # 95 A = [[9.784608]]
Loss = 1.9176414
Step # 100 A = [[9.849552]]
Loss = 1.1561565

四、绘制图像

plt.plot(range(0, 100, 5), loss_stochastic, 'b-', label='Stochastic Loss')
plt.plot(range(0, 100, 5), loss_batch, 'r--', label='Batch Loss, size=20')
plt.legend(loc='upper right', prop={'size': 11})
plt.show()