TensorBoard

Components Of TensorBoard

It is made up of five parts, which are as follows:

1. Scalars:  Given the enormous number of epochs, it's tough to look at the accuracy and inaccuracy for each one, and there's a danger of becoming caught in local minima rather than global ones. This section can be used to tackle these two difficulties. It shows the accuracy and error graphs concerning the epoch.
   
   Img_src
2. Graphs: The "model.summary()" findings are visualised in this section. In other words, it improves the aesthetics of neural network architecture. This makes the process of comprehending architecture much easier.
   
   Img_src
3. Distributions: Neural networks have a lot of layers, and each layer is made up of a lot of biases and weights. The distribution of these hyperparameters is depicted in this section.
   
   Img_src
4. Histograms: Consists of the histogram of these hyperparameters.
   
   Img_src
5. Time-Series: Consists of the values for the same over time. These sections are important for analyzing the hyperparameters' trends and controlling them.

Installing TensorBoard

We can download the tensorboard using the pip command given below:

pip install tensorboard

You can also try this code with Online Python Compiler

Run Code

Starting TensorBoard

To begin, we must first launch the TensorBoard service. To do so, type the commands below into the command prompt. The –logdir argument specifies where TensorBoard data will be saved for visualization. The directory name is 'logs' in this case.

tensorboard --logdir logs

You can also try this code with Online Python Compiler

Run Code

As illustrated below, this will start the TensorBoard service on the default port 6066. http://localhost:6006/ is the URL for the TesnorBoard dashboard.

In the Jupyter notebook, you can issue the following command in the cell.

%tensorboard --logdir logs

You can also try this code with Online Python Compiler

Run Code

Implementation

Let us take that we are working on a dataset and have already compiled our model. Now, how do we visualize the training process using tensorBoard? For that, we call the callback function.

A callback function records the model and its results (error, accuracy, bias, weights, and so on) regularly (epochs). This is necessary because the graphs are based on these values at each epoch.

Please note that the parent path for log_dir below should be the same as the logdir value we gave while starting the TensorBoard service in the second step.

tf_callbacks = tf.keras.callbacks.TensorBoard(log_dir = "logs/fit" , histogram_freq = 1)

You can also try this code with Online Python Compiler

Run Code

Finally, we use the fit() function to train the model. We train it for five epochs, and you'll see that we also passed the callback object we built earlier.

model.fit(x_train, y_train, epochs=200, callbacks = tf_callbacks)

You can also try this code with Online Python Compiler

Run Code

Comparing Different Models Using TensorBoard

Creating a decent Neural Network is a difficult task requiring numerous runs to test various parameters. TensorBoard allows you to compare and visualize the performance of all model runs in the dashboard.

We'll construct training logs in several subfolders within the main folder. The sample below will clarify things for you.

We build the TensorBoard Keras callback object in the first run, and the logs are saved in the 'run1' folder within the main logs folder.

tb_callback = tf.keras.callbacks.

You can also try this code with Online Python Compiler

Run Code

TensorBoard(log_dir="logs/run1", histogram_freq=1)
model.fit(X_train, y_train, epochs=5, callbacks=[tb_callback])

You can also try this code with Online Python Compiler

Run Code

The log path for the second run is run2, as seen below.

tb_callback = tf.keras.callbacks.
TensorBoard(log_dir="logs/run2", histogram_freq=1)
model.fit(X_train, y_train, epochs=5, callbacks=[tb_callback])

You can also try this code with Online Python Compiler

Run Code

When we look at the TensorBoard dashboard now, the accuracy and loss graph will provide information for both runs in orange and blue lines.

Img_src

FAQs

1. What am I able to accomplish with TensorBoard?
TensorBoard is a platform that provides the measurements and visualizations required for machine learning. It allows you to keep track of experimental parameters such as loss and accuracy, visualize the model graph, project embeddings to a lower-dimensional space, etc.

2. What is a callback in TensorBoard?
TensorFlow comes with a visualization tool called TensorBoard. This callback logs TensorBoard events, such as Summary charts of metrics, Visualization of a training graph Histograms of activation.

3. Is TensorBoard a real-time platform?
Without Tensorflow installed, Tensorboard does not show real-time updates when training.

4. Is TensorFlow required for TensorBoard?
TensorBoard is a visualization tool included with every TensorFlow installation. We'll use TensorFlow for many things (like training a huge deep neural network), and they can be complex and confusing.

Key Takeaways

Let us brief out the article.

Firstly, we saw the meaning of tensorboard, its various applications, and its different components. Later we saw how to install tensorboard with its starting process. Lastly, we saw the implementation of tensorboard and how it can be compared to different models using tensorboard. That's all from the article. 

Check out this problem - Largest Rectangle in Histogram

I hope you all like this article. Want to learn more about Data Analysis? Here is an excellent course that can guide you in learning. Can also refer to our Machine Learning course.

Happy Learning, Ninjas!