Tensorflow - Tutorial (6) : TensorBoard 可视化工具

1. TensorBoard

为了更方便 TensorFlow 程序的理解、调试与优化，TensorFlow发布了一套叫做 TensorBoard 的可视化工具。可以用 TensorBoard 来展现 TensorFlow 图像，绘制图像生成的定量指标图以及附加数据。TensorBoard可生成以下4类信息:

Event: 展示训练过程中的统计数据（最值，均值等，误差）变化情况

Image: 展示训练过程中记录的图像

Graph: 展示模型的结构

Histogram: 展示训练过程中记录的数据的分布图

本文以 Tensorflow - Tutorial (4) 中的CNN手写数字识别代码为例，对TensorBoard的相关操作进行演示

用 with tf.name_scope() 为变量名划定范围，同一范围的变量在图中属于同一层级，并且通过”name”属性对变量指定名称

with tf.name_scope('inputs'):
X = tf.placeholder("float", [None, 28, 28, 1], name = 'X')
Y = tf.placeholder("float", [None, 10], name = 'Y')

通过下面代码可查看权重w在训练过程中的变化情况，其中”w_1”是生成的图表的名字

tf.histogram_summary("w_1", w)

通过如下语句查看测试集中预测准确率和loss随迭代次数的变化情况

tf.scalar_summary("accuracy", acc_op)
tf.scalar_summary('loss', cost)

通过如下语句将所有的summary进行合并，一起进行SummaryWriter

merged = tf.merge_all_summaries()
writer = tf.train.SummaryWriter("logs/", sess.graph)

每一轮迭代结束后，通过如下语句记录所有的summary

writer.add_summary(summary, i)  # Write summary

代码运行完毕后，logs文件夹下会生成一个包含time stamp的tfevents文件，进入logs文件夹所在目录，在命令行中输入如下命令来启动TensorBoard

tensorboard --inspect --logdir=logs/

用浏览器打开命令行中显示的url地址，若在结果显示上出现问题，可参考:Frequently Asked Questions

2. 代码

#!/usr/bin/env python
#coding:utf-8
import tensorflow as tf
import numpy as np
import input_data

batch_size = 128
test_size = 256

def init_weights(shape, s):
return tf.Variable(tf.random_normal(shape, stddev=0.01), name = s)

def model(X, w, w2, w3, w4, w_o, p_keep_conv, p_keep_hidden):
with tf.name_scope('layer_1'):
l1a = tf.nn.relu(tf.nn.conv2d(X, w,                       # l1a shape=(?, 28, 28, 32)
strides=[1, 1, 1, 1], padding='SAME'))

l1 = tf.nn.max_pool(l1a, ksize=[1, 2, 2, 1],              # l1 shape=(?, 14, 14, 32)
strides=[1, 2, 2, 1], padding='SAME')

l1 = tf.nn.dropout(l1, p_keep_conv)

with tf.name_scope('layer_2'):
l2a = tf.nn.relu(tf.nn.conv2d(l1, w2,                     # l2a shape=(?, 14, 14, 64)
strides=[1, 1, 1, 1], padding='SAME'))
l2 = tf.nn.max_pool(l2a, ksize=[1, 2, 2, 1],              # l2 shape=(?, 7, 7, 64)
strides=[1, 2, 2, 1], padding='SAME')
l2 = tf.nn.dropout(l2, p_keep_conv)

with tf.name_scope('layer_3'):
l3a = tf.nn.relu(tf.nn.conv2d(l2, w3,                     # l3a shape=(?, 7, 7, 128)
strides=[1, 1, 1, 1], padding='SAME'))
l3 = tf.nn.max_pool(l3a, ksize=[1, 2, 2, 1],              # l3 shape=(?, 4, 4, 128)
strides=[1, 2, 2, 1], padding='SAME')

l3 = tf.reshape(l3, [-1, w4.get_shape().as_list()[0]])    # reshape to (?, 2048)
l3 = tf.nn.dropout(l3, p_keep_conv)

with tf.name_scope('layer_4'):
l4 = tf.nn.relu(tf.matmul(l3, w4))
l4 = tf.nn.dropout(l4, p_keep_hidden)

with tf.name_scope('layer_5'):
pyx = tf.matmul(l4, w_o)
return pyx

mnist = input_data.read_data_sets("MNIST_data/", one_hot=True)

trX, trY, teX, teY = mnist.train.images, mnist.train.labels, mnist.test.images, mnist.test.labels
trX = trX.reshape(-1, 28, 28, 1)
teX = teX.reshape(-1, 28, 28, 1)
with tf.name_scope('inputs'):
X = tf.placeholder("float", [None, 28, 28, 1], name = 'X')
Y = tf.placeholder("float", [None, 10], name = 'Y')
p_keep_conv = tf.placeholder("float", name = 'pro_dropout_conv')
p_keep_hidden = tf.placeholder("float", name = 'pro_dropout_hidden')

with tf.name_scope('weights'):
w = init_weights([3, 3, 1, 32],"w_1")
w2 = init_weights([3, 3, 32, 64],"w_2")
w3 = init_weights([3, 3, 64, 128], "w_3")
w4 = init_weights([128 * 4 * 4, 625], "w_4")
w_o = init_weights([625, 10], "w_o")

# Add histogram summaries for weights
tf.histogram_summary("w_1", w)
tf.histogram_summary("w_2", w2)
tf.histogram_summary("w_3", w3)
tf.histogram_summary("w_4", w4)
tf.histogram_summary("w_o", w_o)

py_x = model(X, w, w2, w3, w4, w_o, p_keep_conv, p_keep_hidden)
with tf.name_scope('loss'):
cost = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(py_x, Y))
tf.scalar_summary('loss', cost) # Add scalar summary for cost
with tf.name_scope('train'):
train_op = tf.train.RMSPropOptimizer(0.001, 0.9).minimize(cost)
with tf.name_scope('prediction'):
correct_pred = tf.equal(tf.argmax(Y, 1), tf.argmax(py_x, 1)) # Count correct predictions
acc_op = tf.reduce_mean(tf.cast(correct_pred, "float")) # Cast boolean to float to average
# Add scalar summary for accuracy
tf.scalar_summary("accuracy", acc_op)
# Launch the graph in a session
with tf.Session() as sess:
merged = tf.merge_all_summaries()
writer = tf.train.SummaryWriter("logs/", sess.graph)# create a log writer

tf.initialize_all_variables().run()

for i in range(30):
loss_sum = 0.0
training_batch = zip(range(0, len(trX), batch_size),
range(batch_size, len(trX)+1, batch_size))
for start, end in training_batch:
_, loss_value, summary = sess.run([train_op, cost, merged], feed_dict={X: trX[start:end], Y: trY[start:end],
p_keep_conv: 0.8, p_keep_hidden: 0.5})
loss_sum +=loss_value
print i, loss_sum

summary = sess.run(merged, feed_dict={X: teX, Y: teY, p_keep_conv: 1.0, p_keep_hidden: 1.0})
writer.add_summary(summary, i)  # Write summary

3. 结果

在EVENTS中查看预测准确率和loss随迭代次数的变化情况



在GRAPH中查看CNN模型的结构



在HISTOGRAMS中查看权重的变化情况