TensorBoard Basics

What is TensorBoard?

TensorBoard is a module of Tensorflow that provides a suite of visualisation tools that help in understanding, debugging and optimize the model created in Tensorflow.

Installing TensorBoard

TensorBoard is installed along with Tensorflow. It can also be installed as a standalone software from Standalone TensorBoard

Key Concepts

TensorBoard works by reading summary data from Tensorflow event files which is generated by tensorflow.

Tensorflow creates a computational graph of the neural network model implemented.This model is then stored as a graph using the summary method.The name field in the Tensorflow variables, placeholders, operations, and name_scopes are used to annotate the nodes in the graph. These graphs are written to files and stored in a directory indicated in the summary.FileWriter method.We can select which nodes to be included in the graph using the summary.merge method. In order to select all the nodes to be included in the graph by using the summary.merge_all method.This summary (in the form of a protocol buffer) is added to the event file using the add_summary method.

#Although the model in this code-segment will not converge, 
#this is just to show how tensorflow creates and saves summaries 
#in event files.
import tensorflow as tf
import numpy as np
N = 5
x = tf.placeholder(tf.float32, shape=[N,1],name='X') 
y = tf.placeholder(tf.float32, shape=[N,N],name='y') 
with tf.name_scope('fc_layer'):
    W = tf.Variable(tf.truncated_normal([1,N], stddev=0.1), name='W')
    b = tf.Variable(tf.constant(0.1,shape=[N]),name='b')
    tf.summary.histogram('weights',W) #values of W will be shown as a histogram plot
    tf.summary.histogram('biases',b) #values of W will be shown as a histogram plot
    y_ = tf.matmul(x,W) + b
with tf.name_scope('loss_func'):    
    
    tf.summary.scalar('loss',0.05)


sess = tf.Session()
merged_summary = tf.summary.merge_all()
writer = tf.summary.FileWriter("./graphs/", graph=sess.graph)
sess.run(tf.global_variables_initializer())
s = sess.run(merged_summary,feed_dict={x:np.ones((N,1)),y:np.ones((N,N))})
i = 1  
writer.add_summary(s,i)
```python

## Running TensorBoard

Now that the event files have been created and stored in the directory[./graphs], running tensorboard with logdir pointing to the directory where the event files have been stored will give us a visualization of the model in the web browser.
```python
tensorboard --logdir=./graphs