Evaluating Results with TensorBoard

In the previous chapter, we understood how a neural network works, what the various hyper parameters in a neural network are, and how they can be tweaked further to improve our model's accuracy.

Google offers TensorBoard, a visualization of the model training logs. In this chapter, we show how to use TensorBoard for TensorFlow and Keras. We interpret the visualizations generated by TensorBoard to understand the performance of our models, and also understand the other functionalities in TensorBoard that can help visualize our dataset better.

As discussed in the previous chapter, Keras as a framework is a wrapper on top of either TensorFlow or Theano. The computations that you'll use TensorFlow for, such as training a massive deep neural network, can be complex and confusing. To make it easier to understand, debug, and optimize TensorFlow...

Setting up TensorBoard

In the previous chapter, we understood how Datalab can be set up. Installing TensorBoard in Datalab is as simple as specifying the following code:

Note that we need not make any separate installations for TensorBoard and it comes in prebuilt within the google.datalab.ml package.

Once the package is imported, we need to start TensorBoard by specifying the location of logs that contain the summaries written by the model fitting process.

The tb.start method works as follows:

Note that, in the first step, it checks whether the user is permitted to perform the calculation. Next, it picks up an unused port to open TensorBoard, and finally it starts TensorBoard along with printing the link to open TensorBoard.

We will learn more about writing to logs in the next section.

Overview of summary operations

Summaries provide a way to export condensed information about a model, which is then accessible in tools such as TensorBoard.

Some of the commonly used summary functions are:

• scalar
• histogram
• audio
• image
• merge
• merge_all

A scalar summary operation returns a scalar, that is, the value of a certain metric over an increasing number of epochs.

A histogram summary operation returns the histogram of various values—potentially weights and biases at each layer.

The image and audio summary operations return images and audio, which can be visualized and played in TensorBoard respectively.

A merge operation returns the union of all the values of input summaries, while merge_all returns the union of all the summaries contained in the model specification.

A visualization of some of the summaries discussed here will be provided in the next section.

...

Summary

In this chapter, we understood visualizing neural network models in TensorBoard, both from Keras and TensorFlow. We also considered how to visualize the models, distribution of weights, and loss/accuracy metrics in both premade estimators and custom defined estimators. And also the various metrics in neural networks.