這篇文章主要介紹TensorFlow如何實(shí)現(xiàn)隨機(jī)訓(xùn)練和批量訓(xùn)練，文中介紹的非常詳細(xì)，具有一定的參考價(jià)值，感興趣的小伙伴們一定要看完！
TensorFlow更新模型變量。它能一次操作一個(gè)數(shù)據(jù)點(diǎn)，也可以一次操作大量數(shù)據(jù)。一個(gè)訓(xùn)練例子上的操作可能導(dǎo)致比較“古怪”的學(xué)習(xí)過(guò)程，但使用大批量的訓(xùn)練會(huì)造成計(jì)算成本昂貴。到底選用哪種訓(xùn)練類型對(duì)機(jī)器學(xué)習(xí)算法的收斂非常關(guān)鍵。
為了TensorFlow計(jì)算變量梯度來(lái)讓反向傳播工作，我們必須度量一個(gè)或者多個(gè)樣本的損失。
隨機(jī)訓(xùn)練會(huì)一次隨機(jī)抽樣訓(xùn)練數(shù)據(jù)和目標(biāo)數(shù)據(jù)對(duì)完成訓(xùn)練。另外一個(gè)可選項(xiàng)是，一次大批量訓(xùn)練取平均損失來(lái)進(jìn)行梯度計(jì)算，批量訓(xùn)練大小可以一次上擴(kuò)到整個(gè)數(shù)據(jù)集。這里將顯示如何擴(kuò)展前面的回歸算法的例子——使用隨機(jī)訓(xùn)練和批量訓(xùn)練。
批量訓(xùn)練和隨機(jī)訓(xùn)練的不同之處在于它們的優(yōu)化器方法和收斂。
重慶分公司
重慶分公司
# 隨機(jī)訓(xùn)練和批量訓(xùn)練
#----------------------------------
#
# This python function illustrates two different training methods:
# batch and stochastic training. For each model, we will use
# a regression model that predicts one model variable.
import matplotlib.pyplot as plt
import numpy as np
import tensorflow as tf
from tensorflow.python.framework import ops
ops.reset_default_graph()
# 隨機(jī)訓(xùn)練:
# Create graph
sess = tf.Session()
# 聲明數(shù)據(jù)
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)
# 聲明變量 (one model parameter = A)
A = tf.Variable(tf.random_normal(shape=[1]))
# 增加操作到圖
my_output = tf.multiply(x_data, A)
# 增加L2損失函數(shù)
loss = tf.square(my_output - y_target)
# 初始化變量
init = tf.global_variables_initializer()
sess.run(init)
# 聲明優(yōu)化器
my_opt = tf.train.GradientDescentOptimizer(0.02)
train_step = my_opt.minimize(loss)
loss_stochastic = []
# 運(yùn)行迭代
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)
# 批量訓(xùn)練:
# 重置計(jì)算圖
ops.reset_default_graph()
sess = tf.Session()
# 聲明批量大小
# 批量大小是指通過(guò)計(jì)算圖一次傳入多少訓(xùn)練數(shù)據(jù)
batch_size = 20
# 聲明模型的數(shù)據(jù)、占位符
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)
# 聲明變量 (one model parameter = A)
A = tf.Variable(tf.random_normal(shape=[1,1]))
# 增加矩陣乘法操作（矩陣乘法不滿足交換律）
my_output = tf.matmul(x_data, A)
# 增加損失函數(shù)
# 批量訓(xùn)練時(shí)損失函數(shù)是每個(gè)數(shù)據(jù)點(diǎn)L2損失的平均值
loss = tf.reduce_mean(tf.square(my_output - y_target))
# 初始化變量
init = tf.global_variables_initializer()
sess.run(init)
# 聲明優(yōu)化器
my_opt = tf.train.GradientDescentOptimizer(0.02)
train_step = my_opt.minimize(loss)
loss_batch = []
# 運(yùn)行迭代
for i in range(100):
rand_index = np.random.choice(100, size=batch_size)
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)
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()輸出：
Step #5 A = [ 1.47604525]
Loss = [ 72.55678558]
Step #10 A = [ 3.01128507]
Loss = [ 48.22986221]
Step #15 A = [ 4.27042341]
Loss = [ 28.97912598]
Step #20 A = [ 5.2984333]
Loss = [ 16.44779968]
Step #25 A = [ 6.17473984]
Loss = [ 16.373312]
Step #30 A = [ 6.89866304]
Loss = [ 11.71054649]
Step #35 A = [ 7.39849901]
Loss = [ 6.42773056]
Step #40 A = [ 7.84618378]
Loss = [ 5.92940331]
Step #45 A = [ 8.15709782]
Loss = [ 0.2142024]
Step #50 A = [ 8.54818344]
Loss = [ 7.11651039]
Step #55 A = [ 8.82354641]
Loss = [ 1.47823763]
Step #60 A = [ 9.07896614]
Loss = [ 3.08244276]
Step #65 A = [ 9.24868107]
Loss = [ 0.01143846]
Step #70 A = [ 9.36772251]
Loss = [ 2.10078788]
Step #75 A = [ 9.49171734]
Loss = [ 3.90913701]
Step #80 A = [ 9.6622715]
Loss = [ 4.80727625]
Step #85 A = [ 9.73786926]
Loss = [ 0.39915398]
Step #90 A = [ 9.81853104]
Loss = [ 0.14876099]
Step #95 A = [ 9.90371323]
Loss = [ 0.01657014]
Step #100 A = [ 9.86669159]
Loss = [ 0.444787]
Step #5 A = [[ 2.34371352]]
Loss = 58.766
Step #10 A = [[ 3.74766445]]
Loss = 38.4875
Step #15 A = [[ 4.88928795]]
Loss = 27.5632
Step #20 A = [[ 5.82038736]]
Loss = 17.9523
Step #25 A = [[ 6.58999157]]
Loss = 13.3245
Step #30 A = [[ 7.20851326]]
Loss = 8.68099
Step #35 A = [[ 7.71694899]]
Loss = 4.60659
Step #40 A = [[ 8.1296711]]
Loss = 4.70107
Step #45 A = [[ 8.47107315]]
Loss = 3.28318
Step #50 A = [[ 8.74283409]]
Loss = 1.99057
Step #55 A = [[ 8.98811722]]
Loss = 2.66906
Step #60 A = [[ 9.18062305]]
Loss = 3.26207
Step #65 A = [[ 9.31655025]]
Loss = 2.55459
Step #70 A = [[ 9.43130589]]
Loss = 1.95839
Step #75 A = [[ 9.55670166]]
Loss = 1.46504
Step #80 A = [[ 9.6354847]]
Loss = 1.49021
Step #85 A = [[ 9.73470974]]
Loss = 1.53289
Step #90 A = [[ 9.77956581]]
Loss = 1.52173
Step #95 A = [[ 9.83666706]]
Loss = 0.819207
Step #100 A = [[ 9.85569191]]
Loss = 1.2197
| 訓(xùn)練類型 | 優(yōu)點(diǎn) | 缺點(diǎn) |
|---|---|---|
| 隨機(jī)訓(xùn)練 | 脫離局部最小 | 一般需更多次迭代才收斂 |
| 批量訓(xùn)練 | 快速得到最小損失 | 耗費(fèi)更多計(jì)算資源 |
以上是“TensorFlow如何實(shí)現(xiàn)隨機(jī)訓(xùn)練和批量訓(xùn)練”這篇文章的所有內(nèi)容，感謝各位的閱讀！希望分享的內(nèi)容對(duì)大家有幫助，更多相關(guān)知識(shí)，歡迎關(guān)注創(chuàng)新互聯(lián)行業(yè)資訊頻道！
