mnist_cnn_dropout_tensorboard.py

# coding: utf-8

# ## 使用mnist数据集，在mnist_cnn_tensorboard的基础上，使用dropout 提高准确率
# 在图模型的全连接层的激活函数前 添加dropout    
# 在训练阶段keep_prob = 0.8,在测试阶段keep_prob =1.0   
# 在一个图上看训练和测试的loss和accuracy  
# 
# 从结果对比可以看出，使用dropout对过拟合有一定效果，但对准确率提高没有作用
# 
# ## 当前准确率为 98.12%
# 
# 如何将两个指标在同一个图上显示  比如 train上的loss和test上的loss；train上的accuracy和test上的accuracy  
# 

# In[1]:

import tensorflow as tf
import numpy as np
import os
print("current work path:",os.getcwd())


# In[2]:

from tensorflow.examples.tutorials.mnist import input_data
mnist = input_data.read_data_sets(".",one_hot=True)


# In[3]:

# 定义图模型

def get_weights(shape,name="weights"):
    w = tf.get_variable(name=name,shape=shape,                    initializer=tf.truncated_normal_initializer(dtype=tf.float32,mean=0,stddev=0.1))
    return w

def get_biases(shape,name="biases"):
    b = tf.get_variable(name=name,shape=shape,initializer=tf.constant_initializer(0.1)) 
    return b

def conv2d(inputs,w,mode,name):
    outputs = tf.nn.conv2d(inputs,w,strides=[1,1,1,1],padding=mode,name=name)
    return outputs

def max_pool(inputs,name):
    outputs = tf.nn.max_pool(inputs,ksize=[1,2,2,1],strides=[1,2,2,1],padding='SAME',name=name)
    return outputs

def inference(inputs,keep_prob):

    with tf.name_scope("inputs"):
        outputs = tf.reshape(inputs,[-1,28,28,1]) 
        tf.summary.image('input',outputs,12) # 可视化 输入数据 Image
        
    with tf.name_scope("conv1"):
        w_c1 = get_weights([3,3,1,6],"conv1_weights")
        b_c1 = get_biases([6],'conv1_biases')
        outputs = conv2d(outputs,w_c1,'SAME','conv1')
        outputs = tf.nn.bias_add(outputs,b_c1)
        outputs = tf.nn.relu(outputs,name='relu1')
    
    with tf.name_scope("pool1"):
        outputs = max_pool(outputs,'pool1')
        
    with tf.name_scope("conv2"):
        w_c2 = get_weights([3,3,6,16],"conv2_weights")
        b_c2 = get_biases([16],'conv2_biases')
        outputs = conv2d(outputs,w_c2,'SAME','conv2')
        outputs = tf.nn.bias_add(outputs,b_c2)
        outputs = tf.nn.relu(outputs,name='relu2')

    with tf.name_scope("pool2"):
        outputs = max_pool(outputs,'pool2')
        
    with tf.name_scope("conv3"):
        w_c3 = get_weights([3,3,16,64],'conv3_weights')
        b_c3 = get_biases([64],'conv3_biases')
        outputs = conv2d(outputs,w_c3,'SAME','conv3')
        outputs = tf.nn.bias_add(outputs,b_c3)
        outputs = tf.nn.relu(outputs,name='relu3')

    with tf.name_scope("flatten"):
        outputs = tf.reshape(outputs,shape=[-1,7*7*64])
        
    with tf.name_scope("fc1"):
        w_fc = get_weights([7*7*64,84],"fc_weights")
        b_fc = get_biases([84],"fc_biases")
        outputs = tf.matmul(outputs,w_fc) + b_fc
        
        # 在全连接层的激活函数前使用dropout，可以减少过拟合
        outputs = tf.nn.dropout(keep_prob=keep_prob,x=outputs)
        
        outputs = tf.nn.relu(outputs,name="fc1")

    with tf.name_scope("softmax"):
        w_softmax = get_weights([84,10],"softmax_weights")
        b_softmax = get_biases([10],"softmax_biases")
#         index_summary_visualization(index=w_softmax,name="weights") #可视化监控指标 w_softmax
        tf.summary.histogram("w_softmax",w_softmax) #可视化监控指标 w_softmax
#         index_summary_visualization(index=b_softmax,name="biases") #可视化监控指标 b_softmax
        tf.summary.histogram("b_softmax",b_softmax) #可视化监控指标 b_softmax
        outputs = tf.matmul(outputs,w_softmax) + b_softmax
        outputs = tf.nn.softmax(outputs,name="softmax")
    
    return outputs


# In[4]:

# 定义计算模型

learning_rate = 0.01


# with tf.name_scope("inputs"):
x = tf.placeholder(dtype=tf.float32,shape=[None,784],name='image')
y = tf.placeholder(dtype=tf.float32,shape=[None,10],name='label')
keep_prob =tf.placeholder(dtype=tf.float32,shape=[],name="keep_prob")
y_predict = inference(x,keep_prob)

with tf.name_scope("loss"):
    loss = tf.reduce_mean(tf.reduce_sum(-y*tf.log(y_predict+1e-8),reduction_indices=[1]))
#     index_summary_visualization(loss,"loss")
    tf.summary.scalar("cross_entropy",loss) # 可视化监控指标 loss
    

with tf.name_scope("optimizer"):
    optimizer = tf.train.AdamOptimizer(learning_rate)
    train_step = optimizer.minimize(loss)
with tf.name_scope("result_index"):
    correct_prediction = tf.equal(tf.argmax(y,1),tf.argmax(y_predict,1)) # 正确数
    accuracy = tf.reduce_mean(tf.cast(correct_prediction,tf.float32)) # 正确率 
    tf.summary.scalar("accuracy",accuracy)
    
# 能不能在同一个图上显示两个scalar
# with tf.name_scope("scalar"):
#     tf.summary.scalar(["cross_entropy","accuracy"],[loss,accuracy]) # 可视化监控指标 loss
    


# In[5]:

# 定义会话

workspace_path =  "E://home//workspace//python//" #"D://"
log_path = "log"
training_steps = 5001
batch_size = 128


with tf.Session() as sess:
    # 生成一个写日志的writer，并将当前的TensorFlow 计算题写入日志。
    writer = tf.summary.FileWriter(workspace_path+log_path,sess.graph) # 可视化计算图
    summary_merged = tf.summary.merge_all()
    sess.run(tf.global_variables_initializer())
    
         
    for step in range(training_steps):
        xs,ys = mnist.train.next_batch(batch_size)
        sess.run(train_step,feed_dict={x:xs,y:ys,keep_prob:0.5})
        
        if step%100 == 0:
            summary,validation_loss,validation_acc = sess.run            ([summary_merged,loss,accuracy],{x:mnist.validation.images,y:mnist.validation.labels,keep_prob:1.0})
            writer.add_summary(summary,global_step=step) ### 将可视化指标写入日志文件
            print("after %d steps,loss is %.2f,and accuracy is %.2f%% on Validation data."%(step,validation_loss,validation_acc*100))
            
        if step%1000 ==0:
            test_loss,test_acc = sess.run([loss,accuracy],{x:mnist.test.images,y:mnist.test.labels,keep_prob:1.0})
            print("\n::: After %d steps,loss is %.2f,and accuracy is %.2f%% on Test data. "%(step,test_loss,test_acc*100))
            train_loss,train_acc = sess.run([loss,accuracy],{x:xs,y:ys,keep_prob:1.0})
            print("::: After %d steps,loss is %.2f,and accuracy is %.2f%% on Train data. \n"%(step,train_loss,train_acc*100))
    print("log save path：",workspace_path + log_path)
    # 关闭写日志的writer        
    writer.close()