tensorflow学习（11）使用tensorboard可视化手写数字识别训练过程

上一节可视化了训练的网络的权值偏置值等数据的曲线图&＃xff0c;这节可视化网络的训练过程。

import tensorflow as tf
from tensorflow.examples.tutorials.mnist import input_data
from tensorflow.contrib.tensorboard.plugins import projector#读取数据集
mnist &＃61; input_data.read_data_sets("../MNIST_DATA", one_hot&＃61;True)
max_steps&＃61;1001#运行次数
image_num&＃61;3000#图片数量
DIR&＃61;&＃39;E:/py_code&＃39;#文件路径sess&＃61;tf.Session()#定义会话#载入图片&＃xff08;将3000个图片打包成一个矩阵&＃xff09;
embedding&＃61;tf.Variable(tf.stack(mnist.test.images[:image_num]),trainable&＃61;False,name&＃61;&＃39;embedding&＃39;)#参数概要
def variable_summaries(var): # 在tensorboard中显示var的相关属性值with tf.name_scope("summaries"):mean &＃61; tf.reduce_mean(var) # 平均值tf.summary.scalar("mean", mean)with tf.name_scope("stddev"):stddev &＃61; tf.sqrt(tf.reduce_mean(tf.square(var - mean)))tf.summary.scalar("stddev", stddev) # 标准差tf.summary.scalar("max", tf.reduce_max(var)) # 最大值tf.summary.scalar("min", tf.reduce_min(var)) # 最小值tf.summary.histogram("histogram", var) # 直方图
#命名空间
with tf.name_scope(&＃39;input&＃39;):x&＃61;tf.placeholder(tf.float32,[None,784],name&＃61;&＃39;x_input&＃39;)y&＃61;tf.placeholder(tf.float32,[None,10],name&＃61;&＃39;y_input&＃39;)
#显示10个数字的图片
with tf.name_scope(&＃39;input_reshape&＃39;):image_shaped_input&＃61;tf.reshape(x,[-1,28,28,1])tf.summary.image(&＃39;input&＃39;,image_shaped_input,10)
#创建简单神经网络
with tf.name_scope(&＃39;layer&＃39;):with tf.name_scope(&＃39;weights&＃39;):w&＃61;tf.Variable(tf.random_normal([784,10]),name&＃61;&＃39;W&＃39;)variable_summaries(w)with tf.name_scope(&＃39;biases&＃39;):b&＃61;tf.Variable(tf.zeros([10]),&＃39;b&＃39;)variable_summaries(b)with tf.name_scope(&＃39;Wx_plus_b&＃39;):#线性Wx_plus_b&＃61;tf.matmul(x,w)&＃43;bwith tf.name_scope(&＃39;softmax&＃39;):#通过softmax激活函数变成非线性y_pred&＃61;tf.nn.softmax(Wx_plus_b)
#交叉熵代价函数
with tf.name_scope(&＃39;loss&＃39;):loss&＃61;tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits_v2(labels&＃61;y,logits&＃61;y_pred))tf.summary.scalar(&＃39;loss&＃39;,loss)
#精确度
with tf.name_scope(&＃39;accuracy&＃39;):accuracy&＃61;tf.reduce_mean(tf.cast(tf.equal(tf.argmax(y,1),tf.argmax(y_pred,1)),tf.float32))# argmax返回一维张量中最大值所在的位置tf.summary.scalar(&＃39;accuracy&＃39;,accuracy)
#使用梯度下降算法来训练&＃xff0c;使得loss最小&＃xff08;可优化&＃xff09;
with tf.name_scope(&＃39;train&＃39;):train_step&＃61;tf.train.GradientDescentOptimizer(0.2).minimize(loss)#初始化变量
sess.run(tf.global_variables_initializer())
#产生metadata文件
if tf.gfile.Exists(DIR &＃43; &＃39;/projector/projector/metadata.tsv&＃39;):tf.gfile.DeleteRecursively(DIR &＃43; &＃39;/projector/projector/metadata.tsv&＃39;)
with open(DIR &＃43; &＃39;/projector/projector/metadata.tsv&＃39;, &＃39;w&＃39;) as f:labels &＃61; sess.run(tf.argmax(mnist.test.labels[:], 1))for i in range(image_num):f.write(str(labels[i]) &＃43; &＃39;\n&＃39;)merged &＃61; tf.summary.merge_all()projector_writer &＃61; tf.summary.FileWriter(DIR &＃43; &＃39;/projector/projector&＃39;, sess.graph)#定义一个writer&＃xff0c;写入图
# writer&＃61;tf.summary.FileWriter(DIR&＃43;&＃39;/projector/projector/logs&＃39;,sess.graph)
saver &＃61; tf.train.Saver()#保存图
config &＃61; projector.ProjectorConfig()#配置
embed &＃61; config.embeddings.add()
embed.tensor_name &＃61; embedding.name
embed.metadata_path &＃61; DIR &＃43; &＃39;/projector/projector/metadata.tsv&＃39;
embed.sprite.image_path &＃61; DIR &＃43; &＃39;/mnist_10k_sprite.png&＃39;
embed.sprite.single_image_dim.extend([28, 28])
projector.visualize_embeddings(projector_writer, config)#开始训练
for i in range(max_steps):batch_xs, batch_ys &＃61; mnist.train.next_batch(100)run_options &＃61; tf.RunOptions(trace_level&＃61;tf.RunOptions.FULL_TRACE)run_metadata &＃61; tf.RunMetadata()summary, _ &＃61; sess.run([merged, train_step], feed_dict&＃61;{x: batch_xs, y: batch_ys}, options&＃61;run_options,run_metadata&＃61;run_metadata)projector_writer.add_run_metadata(run_metadata, &＃39;step%03d&＃39; % i)projector_writer.add_summary(summary, i)if i % 100 &＃61;&＃61; 0:#每100个图片打印一次结果acc &＃61; sess.run(accuracy, feed_dict&＃61;{x: mnist.test.images, y: mnist.test.labels})print(&＃39;Iter &＃39; &＃43; str(i) &＃43; &＃39;, accuracy &＃39; &＃43; str(acc))saver.save(sess, DIR &＃43; &＃39;/projector/projector/a_model.ckpt&＃39;, global_step&＃61;max_steps)
projector_writer.close()
sess.close()


1、首先要使用这个代码需要修改这个文件路径&＃xff1a;

DIR&＃61;&＃39;E:/py_code&＃39;#文件路径


2、然后在文件路劲下创建两级projector子目录&＃xff08;不然会报错找不到路径&＃xff0c;但是网课说的是会自动创建至于为什么不行也就没有深究&＃xff09;&＃xff1a;

3、修改你的图片路径&＃xff1a;

embed.sprite.image_path &＃61; DIR &＃43; &＃39;/mnist_10k_sprite.png&＃39;


我这里就是直接放在py_code底下了&＃xff0c;图片有1.5M。。。csdn上传图片有水印所以就放在csdn的资源里吧。
下载链接&＃xff1a;训练图片下载传送门

1、程序运行后可以在projector子目录下看到生成的文件&＃xff1a;

2、win &＃43; R打开命令运行窗口&＃xff0c;输入cmd回车打开shell&＃xff1a;

3、切换到文件对应的盘&＃xff1a;

e:


4、运行tensorboard可视化&＃xff1a;

tensorboard --logdir&＃61;E:\py_code\tensorflow\logs


5、复制最底下的url&＃xff1a;http://DESKTOP-V48600Q:6006在浏览器打开&＃xff0c;用localhost加端口号6006也能打开。
6、切换到EMBEDDINGS栏如下图&＃xff1a;

7、选择T-SNE:

8、点击左侧中间部位的Re-run可以重新运行模型&＃xff0c;将Color by一栏选择lable 10 colors可以区分数字的颜色&＃xff1a;

9、训练400次的结果&＃xff1a;

10、修改优化器提高准确度&＃xff1a;

import tensorflow as tf
from tensorflow.examples.tutorials.mnist import input_data
from tensorflow.contrib.tensorboard.plugins import projector#读取数据集
mnist &＃61; input_data.read_data_sets("../MNIST_DATA", one_hot&＃61;True)
max_steps &＃61; 1001 # 运行次数
image_num &＃61; 3000 # 图片数量
DIR &＃61; &＃39;E:/py_code&＃39; # 文件路径sess &＃61; tf.Session() # 定义会话# 载入图片&＃xff08;将3000个图片打包成一个矩阵&＃xff09;
embedding&＃61;tf.Variable(tf.stack(mnist.test.images[:image_num]),trainable&＃61;False,name&＃61;&＃39;embedding&＃39;)# 命名空间
x &＃61; tf.placeholder(tf.float32, [None, 784], name&＃61;&＃39;x_input&＃39;)
y &＃61; tf.placeholder(tf.float32, [None, 10], name&＃61;&＃39;y_input&＃39;)
# dropout的所要训练神经元的比例
keep_prob &＃61; tf.placeholder(tf.float32)
# 学习率
lr &＃61; tf.Variable(0.001, dtype&＃61;tf.float32)# 显示10个数字的图片
image_shaped_input&＃61;tf.reshape(x, [-1, 28, 28, 1])
tf.summary.image(&＃39;input&＃39;, image_shaped_input, 10)# 创建简单神经网络
W1 &＃61; tf.Variable(tf.truncated_normal([784, 500], stddev&＃61;0.1))
b1 &＃61; tf.Variable(tf.zeros([500]) &＃43; 0.1)
L1 &＃61; tf.nn.relu(tf.matmul(x, W1) &＃43; b1)W2 &＃61; tf.Variable(tf.truncated_normal([500, 300], stddev&＃61;0.1))
b2 &＃61; tf.Variable(tf.zeros([300]) &＃43; 0.1)
L2 &＃61; tf.nn.relu(tf.matmul(L1, W2) &＃43; b2)W4 &＃61; tf.Variable(tf.truncated_normal([300, 10], stddev&＃61;0.1))
b4 &＃61; tf.Variable(tf.zeros([10]) &＃43; 0.1)
y_pred &＃61; tf.nn.softmax(tf.matmul(L2, W4) &＃43; b4)# 交叉熵代价函数
loss&＃61;tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits_v2(labels&＃61;y,logits&＃61;y_pred))
# 精确度
accuracy&＃61;tf.reduce_mean(tf.cast(tf.equal(tf.argmax(y,1),tf.argmax(y_pred,1)),tf.float32)) # argmax返回一维张量中最大值所在的位置
# 使用梯度下降法训练&＃xff0c;使得loss最小&＃xff08;#可优化&＃xff09;
train_step &＃61; tf.train.AdamOptimizer(lr).minimize(loss)# 初始化变量
sess.run(tf.global_variables_initializer())
# 产生metadata文件
if tf.gfile.Exists(DIR &＃43; &＃39;/projector/projector/metadata.tsv&＃39;):tf.gfile.DeleteRecursively(DIR &＃43; &＃39;/projector/projector/metadata.tsv&＃39;)
with open(DIR &＃43; &＃39;/projector/projector/metadata.tsv&＃39;, &＃39;w&＃39;) as f:labels &＃61; sess.run(tf.argmax(mnist.test.labels[:], 1))for i in range(image_num):f.write(str(labels[i]) &＃43; &＃39;\n&＃39;)projector_writer &＃61; tf.summary.FileWriter(DIR &＃43; &＃39;/projector/projector&＃39;, sess.graph) # 定义一个writer&＃xff0c;写入图
# writer&＃61;tf.summary.FileWriter(DIR&＃43;&＃39;/projector/projector/logs&＃39;,sess.graph)
saver &＃61; tf.train.Saver() # 保存图
config &＃61; projector.ProjectorConfig() # 配置
embed &＃61; config.embeddings.add()
embed.tensor_name &＃61; embedding.name
embed.metadata_path &＃61; DIR &＃43; &＃39;/projector/projector/metadata.tsv&＃39;
embed.sprite.image_path &＃61; DIR &＃43; &＃39;/mnist_10k_sprite.png&＃39;
embed.sprite.single_image_dim.extend([28, 28])
projector.visualize_embeddings(projector_writer, config)# 开始训练
for i in range(max_steps):sess.run(tf.assign(lr, 0.001 * (0.95 ** i)))batch_xs, batch_ys &＃61; mnist.train.next_batch(100)run_options &＃61; tf.RunOptions(trace_level&＃61;tf.RunOptions.FULL_TRACE)run_metadata &＃61; tf.RunMetadata()summary &＃61; sess.run(train_step, feed_dict&＃61;{x: batch_xs, y: batch_ys}, options&＃61;run_options,run_metadata&＃61;run_metadata)projector_writer.add_run_metadata(run_metadata, &＃39;step%03d&＃39; % i)if i % 100 &＃61;&＃61; 0: # 每100个图片打印一次结果acc &＃61; sess.run(accuracy, feed_dict&＃61;{x: mnist.test.images, y: mnist.test.labels})print(&＃39;Iter &＃39; &＃43; str(i) &＃43; &＃39;, accuracy &＃39; &＃43; str(acc))saver.save(sess, DIR &＃43; &＃39;/projector/projector/a_model.ckpt&＃39;, global_step&＃61;max_steps)
projector_writer.close()
sess.close()


这里注意下再次运行代码的时候需要把这些文件全删了&＃xff08;不然会运行报错&＃xff09;&＃xff1a;