Getting Started with TensorBoard: Visualizing a Simple Logistic Regression in TensorFlow

TensorBoard can display the full computational graph of a neural network as well as various trend curves such as histograms and scalars.

The article walks through a simple logistic‑regression experiment to demonstrate how to use TensorBoard for visualization. The procedure consists of the following steps:

Wrap variables with tf.variable_scope('custom_label') to group them.

Create histogram summaries with tf.summary.histogram('custom_hist', variable_to_monitor) .

Create scalar summaries with tf.summary.scalar('custom_scalar', scalar_to_monitor) .

Merge all summaries using tf.summary.merge_all() , which returns a handle merge_op for later execution.

Instantiate a FileWriter to write logs to a local directory.

During training, call FileWriter.add_summary(result, step) to write each step’s values into the TensorBoard buffer.

The complete Python code (with detailed comments) is shown below.

# -*- coding: utf-8 -*-
# -*- author: knock -*-

import tensorflow as tf
import numpy as np
'''TensorBoard: Visualize logistic‑regression training process'''
# 1. Generate synthetic data
x = np.linspace(-1, 1, 100)[:, np.newaxis]  # shape (100, 1)
noise = np.random.normal(0, 0.1, size=x.shape)  # noise
y = np.power(x, 2) + noise  # target values
# 2. Define TensorFlow placeholders
with tf.variable_scope('Inputs'):
    input_x = tf.placeholder(tf.float32, x.shape, name='input_x')
    input_y = tf.placeholder(tf.float32, y.shape, name='input_y')
# 3. Build a simple network: input → hidden (ReLU) → output
with tf.variable_scope('Net'):
    hidden_1 = tf.layers.dense(input_x, 10, tf.nn.relu, name='hidden_layer')
    output = tf.layers.dense(hidden_1, 1, name='output_layer')
    # Add histograms for hidden layer and predictions
    tf.summary.histogram('hidden_out', hidden_1)
    tf.summary.histogram('prediction', output)
# 4. Compute mean‑squared error loss
loss = tf.losses.mean_squared_error(input_y, output, scope='loss')
# 5. Optimize loss with gradient descent
train_op = tf.train.GradientDescentOptimizer(learning_rate=0.5).minimize(loss)
# Record loss as a scalar summary
tf.summary.scalar('loss', loss)

sess = tf.Session()
sess.run(tf.global_variables_initializer())

writer = tf.summary.FileWriter('./logs', sess.graph)  # write graph and summaries
merge_op = tf.summary.merge_all()

# 6. Train for 100 steps and write summaries each step
for step in range(100):
    _, result = sess.run([train_op, merge_op], {input_x: x, input_y: y})
    writer.add_summary(result, step)

# To launch TensorBoard, run in a terminal:
# tensorboard --logdir logs
# Then open http://localhost:6006 in a browser.

Running the command tensorboard --logdir logs starts TensorBoard (e.g., at http://0.0.0.0:6006), where the user can explore the computational graph, histogram distributions of hidden activations and predictions, and the loss curve over training steps.