Building a Basic Keras Neural Network Sequential Model
The approach basically coincides with Chollet's Keras 4 step workflow, which he outlines in his book "Deep Learning with Python," using the MNIST dataset, and the model built is a Sequential network of Dense layers. A building block for additional posts.
As the title suggest, this post approaches building a basic Keras neural network using the Sequential model API. The specific task herein is a common one (training a classifier on the MNIST dataset), but this can be considered an example of a template for approaching any such similar task.
The approach basically coincides with Chollet's Keras 4 step workflow, which he outlines in his book "Deep Learning with Python," and really amounts to little more than what can be found as an example in the early chapters of the book, or the official Keras tutorials.
The motivation for producing such a post is to use it is a foundational reference for a series of upcoming posts configuring Keras in a variety of different ways. This seemed a better idea than covering the same things over and over again at the start of each post. The content should be useful on its own for those who do not have experience approaching building a neural network in Keras.
Image taken from screenshot of the Keras documentation website
The dataset used is MNIST, and the model built is a Sequential network of Dense layers, intentionally avoiding CNNs for now.
First are the imports and a few hyperparameter and data resizing variables.
Next is a function for outputting some simple (but useful) metadata of our dataset. Since we will be using it a few times, it makes sense to put the few lines in a callable function. Reusable code is an end in and of itself :)
Next we load our dataset (MNIST, using Keras' dataset utilities), and then use the function above to get some dataset metadata.
To feed MNIST instances into a neural network, they need to be reshaped, from a 2 dimensional image representation to a single dimension sequence. We also convert our class vector to a binary matrix (using
to_categorical). This is accomplished below, after which the same function defined above is called again in order to show the effects of our data reshaping.
Both of the required data transformations have been accomplished. Now it's time to build, compile, and train a neural network. You can see more about this process in this previous post.
The only unorthodox (as far as using the Keras library standalone) step has been the use of the Live Loss Plot callback which outputs epoch-by-epoch loss functions and accuracies at the end of each epoch of training. Make sure you have installed Live Loss Plot prior to running the above code. We are also given the final loss and accuracy on our test dataset.
Almost done, but first let's output a summary of the neural network we built.
And finally, visualize the model:
The full code is shown below:
As previously stated, this post doesn't cover anything innovative, but we will publish a series of upcoming posts using Keras which hopefully will be more interesting to the reader, and this common starting point should be beneficial for reference.
Also, for those looking for a streamlined approach to building neural networks using the Keras Sequential model, this post should serve as a basic guide to hitting all the important points along the way. What you do after training is up to you (at this point), but we will circle back around to this in the future as well.
- The Keras 4 Step Workflow
- 7 Steps to Mastering Deep Learning with Keras
- Today I Built a Neural Network During My Lunch Break with Keras
Top Stories Past 30 Days