Tensorflow入门系列（一）

第一个项目

第一个项目是使用简单的多层神经网络（不涉及CNN）对手写字进行的识别，准确率达到98%以上。这个项目主要是初步认识如何进行参数调优的过程。首先贴出实现的程序，然后再来谈谈调优过程遇到的问题。

程序

from __future__ import absolute_import
from __future__ import division
from __future__ import print_function

import argparse
import sys

from tensorflow.examples.tutorials.mnist import input_data

import tensorflow as tf 
FLAGS=None

data_dir='/tmp/tensorflow/mnist/input_data'
mnist= input_data.read_data_sets(data_dir,one_hot=True)

hd1input=784

hd1output=500
hd2output=300
hd3output=10


#输入
x=tf.placeholder(tf.float32,[None,hd1input])

#实际标签
y_=tf.placeholder(tf.float32,[None,10])

#定义权重w和偏差项b
def init_weight(shape,st_dev):
    '''
    random_normal用于生成正太分布随机数，
    均值mean,标准差stddev，shape为矩阵大小
    '''
    weight=tf.Variable(tf.random_normal(shape,stddev=st_dev))
    return(weight)

def init_bias(shape,st_dev):
    bias=tf.Variable(tf.random_normal(shape,stddev=st_dev))
    return(bias)

#定义全连接层
def full_connect(input_layer,weight,bias):
    layer=tf.add(tf.matmul(input_layer,weight),bias)
    return (tf.nn.relu(layer))

#dropout
keep_prob = tf.placeholder(tf.float32) 
#第一层

w1=init_weight([hd1input,hd1output],0.1)
b1=init_bias([hd1output],0.1)

hidden_1=full_connect(x,w1,b1)
L1_drop=tf.nn.dropout(hidden_1,keep_prob) 
#第二层
w2=init_weight([hd1output,hd2output],0.1)
b2=init_bias([hd2output],0.1)
hidden_2=full_connect(L1_drop,w2,b2)

L2_drop=tf.nn.dropout(hidden_2,keep_prob) 
#输出层

w3=init_weight([hd2output,hd3output],0.1)
b3=init_weight([hd3output],0.1)

y=(tf.add(tf.matmul(L2_drop,w3),b3))

#计算交叉熵
cross_entropy = tf.reduce_mean(
    tf.nn.softmax_cross_entropy_with_logits(labels=y_,logits=y)
)

#学习率
learn_rate=tf.Variable(0.001)
train_step = tf.train.AdamOptimizer(learn_rate).minimize(cross_entropy)
#正确率
correct_prediction = tf.equal(tf.argmax(y,1),tf.argmax(y_,1))
accuracy=tf.reduce_mean(tf.cast(correct_prediction,tf.float32))

#参数初始化
init_op=tf.global_variables_initializer()


with tf.Session() as sess:
    sess.run(init_op)

    #学习率改变
    for epoch in range(50):
        sess.run(tf.assign(learn_rate,0.001 * (0.95 ** epoch)))
        #最多5000次迭代
        for _ in range(5000):
            batch_xs,batch_ys=mnist.train.next_batch(100)
            #在会话中，占位符可以使用 feed_dict 馈送数据。
            sess.run(train_step,feed_dict={x:batch_xs,y_:batch_ys,keep_prob:1.0})

        lr=sess.run(learn_rate)
        acc=sess.run(accuracy,feed_dict={x: mnist.test.images,y_: mnist.test.labels,keep_prob:1.0})
        print(" Iter "+ str(epoch) + ",Test Accuracy= "+ str(acc) + ",Learning_Rate ="+str(lr)+" ")

输出

 Iter 0,Test Accuracy= 0.9799,Learning_Rate =0.001 
 Iter 1,Test Accuracy= 0.9827,Learning_Rate =0.00095 
 Iter 2,Test Accuracy= 0.9782,Learning_Rate =0.0009025 
 Iter 3,Test Accuracy= 0.9818,Learning_Rate =0.000857375 
 Iter 4,Test Accuracy= 0.981,Learning_Rate =0.000814506 
 Iter 5,Test Accuracy= 0.9831,Learning_Rate =0.000773781 
 Iter 6,Test Accuracy= 0.9848,Learning_Rate =0.000735092 
 Iter 7,Test Accuracy= 0.9852,Learning_Rate =0.000698337 
 Iter 8,Test Accuracy= 0.9851,Learning_Rate =0.00066342 
 Iter 9,Test Accuracy= 0.9856,Learning_Rate =0.000630249 
 Iter 10,Test Accuracy= 0.9856,Learning_Rate =0.000598737 
 Iter 11,Test Accuracy= 0.9858,Learning_Rate =0.0005688 
 Iter 12,Test Accuracy= 0.9855,Learning_Rate =0.00054036 
 Iter 13,Test Accuracy= 0.984,Learning_Rate =0.000513342 
 Iter 14,Test Accuracy= 0.9844,Learning_Rate =0.000487675 
 Iter 15,Test Accuracy= 0.9849,Learning_Rate =0.000463291 
 Iter 16,Test Accuracy= 0.9849,Learning_Rate =0.000440127 
 Iter 17,Test Accuracy= 0.985,Learning_Rate =0.00041812 
 Iter 18,Test Accuracy= 0.9847,Learning_Rate =0.000397214
 Iter 19,Test Accuracy= 0.9847,Learning_Rate =0.000377354 
 Iter 20,Test Accuracy= 0.9848,Learning_Rate =0.000358486 
 Iter 21,Test Accuracy= 0.985,Learning_Rate =0.000340562 
 Iter 22,Test Accuracy= 0.9852,Learning_Rate =0.000323534 
 Iter 23,Test Accuracy= 0.9854,Learning_Rate =0.000307357 
 Iter 24,Test Accuracy= 0.9852,Learning_Rate =0.000291989 
 Iter 25,Test Accuracy= 0.9852,Learning_Rate =0.00027739 
 Iter 26,Test Accuracy= 0.985,Learning_Rate =0.00026352 
 Iter 27,Test Accuracy= 0.9846,Learning_Rate =0.000250344 
 Iter 28,Test Accuracy= 0.9846,Learning_Rate =0.000237827 
 Iter 29,Test Accuracy= 0.9838,Learning_Rate =0.000225936 
 Iter 30,Test Accuracy= 0.9838,Learning_Rate =0.000214639 
 Iter 31,Test Accuracy= 0.9849,Learning_Rate =0.000203907 
 Iter 32,Test Accuracy= 0.9851,Learning_Rate =0.000193711
 Iter 33,Test Accuracy= 0.985,Learning_Rate =0.000184026 
 Iter 34,Test Accuracy= 0.985,Learning_Rate =0.000174825 
 Iter 35,Test Accuracy= 0.9852,Learning_Rate =0.000166083 
 Iter 36,Test Accuracy= 0.9849,Learning_Rate =0.000157779 
 Iter 37,Test Accuracy= 0.9845,Learning_Rate =0.00014989 
 Iter 38,Test Accuracy= 0.9843,Learning_Rate =0.000142396 
 Iter 39,Test Accuracy= 0.9843,Learning_Rate =0.000135276 
 Iter 40,Test Accuracy= 0.9848,Learning_Rate =0.000128512 
 Iter 41,Test Accuracy= 0.9845,Learning_Rate =0.000122087 
 Iter 42,Test Accuracy= 0.9843,Learning_Rate =0.000115982 
 Iter 43,Test Accuracy= 0.9844,Learning_Rate =0.000110183 
 Iter 44,Test Accuracy= 0.9846,Learning_Rate =0.000104674 
 Iter 45,Test Accuracy= 0.9845,Learning_Rate =9.94403e-05 
 Iter 46,Test Accuracy= 0.9845,Learning_Rate =9.44682e-05 
 Iter 47,Test Accuracy= 0.9844,Learning_Rate =8.97448e-05 
 Iter 48,Test Accuracy= 0.9841,Learning_Rate =8.52576e-05 
 Iter 49,Test Accuracy= 0.9834,Learning_Rate =8.09947e-05

问题

• 优化器

  一开始我使用的是tf.train.GradientDescentOptimizer这个优化器对w进行梯度下降计算，发现准确率最高也就达到93%多（未对w,b进行优化），而采用tf.train.AdamOptimizer这个优化器对w进行梯度下降计算，发现准确率最高达到97%多（未对w,b进行优化），感觉在这个项目中似乎AdamOptimizer会更好一点。

• w,b的初始化

  采用了AdamOptimizer，最高也就达到了97%，没有满足要求。因此考虑到变量参数初始值的问题。一开始的时候，我选择的是w,b均为均值为0，标准差为1的正态分布。考虑到有可能是初始值过大的问题，便改为均值为0，标准差为0.1的正态分布。最终结果如上面所述。

• 交叉熵的计算

  交叉熵的计算我的是
  y=(tf.add(tf.matmul(L2_drop,w3),b3))
  cross_entropy = tf.reduce_mean(
  tf.nn.softmax_cross_entropy_with_logits(labels=y_,logits=y)
  )
  而在其他的一些博客上我看到的一般是
  prediction = tf.nn.softmax(tf.matmul(L2_drop, W3) + b3)
  loss = tf.reduce_mean(
  tf.nn.softmax_cross_entropy_with_logits(labels=y, logits=prediction)
  )
  我尝试加入tf.nn.softmax函数，但是发现准确率变得贼差，完全不能看。而且，根据我自己的理解softmax_cross_entropy_with_logits这个函数中logits参数应该就是wx+b的结果，不需要进行其他处理的。不知道我理解的对不对。

最后

这是我第一个尝试的项目，可能存在一些错误，希望大家指教！