Defining neural networks with Keras
1. Defining neural networks with Keras
In chapter 3, we saw how to define neural networks in TensorFlow, both using linear algebra and higher level Keras operations. In this lesson, we will introduce the Keras sequential API, and expand on our brief and informal introduction of the Keras functional API.2. Classifying sign language letters
Throughout this chapter, we'll focus on using Keras to classify four letters from the Sign Language MNIST dataset: a, b, c, and d. Note that the images appear to be low resolution because each is represented by a 28x28 matrix.3. The sequential API
Now, let's say we experiment with several different architectures and select the one that makes the most accurate predictions. It has an input layer, a first hidden layer with 16 nodes, and a second hidden layer with 8 nodes. We'll have 4 output nodes, since there are 4 letters in the dataset.4. The sequential API
A good way to construct this model in Keras is to use the sequential API. This API is simpler and makes strong assumptions about how you will construct your model. It assumes that you have an input layer, some number of hidden layers, and an output layer. All of these layers are ordered one after the other in a sequence.5. Building a sequential model
We'll start by importing tensorflow. We can then define a sequential model, which we'll name model. Once we have defined this object, we can simply stack layers on top of it sequentially using the add method. Let's start by adding the first hidden layer, which is a dense layer with 16 nodes. We'll select a relu activation function and supply an input_shape, which Keras requires for the first layer. This input shape is simply a tuple that contains the dimensions of our data. Since we'll be using 28 by 28 pixel images, reshaped into vector, we'll supply 28*28 comma as the input shape.6. Building a sequential model
Next, we'll define a second hidden layer according to the desired model architecture. Finally, we specify that the model has 4 output nodes and uses a softmax activation function. If we want to check our model's architecture, we can use the dot summarize method, which we'll return to in the upcoming exercises. The model has now been defined, but it is not yet ready to be trained. We must first perform a compilation step, where we specify the optimizer and loss function. Here, we've selected the adam optimizer and the categorical crossentropy loss function, which we'll use for classification problems with more than 2 classes.7. The functional API
But what if you want to train two models jointly to predict the same target? The functional API is for that.8. Using the functional API
As an example, let's say we have a set of 28x28 images and a set of 10 features of metadata. We want to use both to predict the image's class, but restrict how they interact in our model. We'll start by using the Keras inputs operation to define the input shapes for model 1 and model 2. Next, we define layer 1 and layer 2 as dense layers for model 1. Note that we have to pass the previous layer as an argument if we use the functional API, but did not with the sequential. You may remember that we did this in chapter 3. We were also using the functional API then.9. Using the functional API
We now define layers 1 and 2 for model 2 and then use the add layer in keras to combine the outputs in a layer that merges the two models. Finally, we define a functional model. As inputs, it takes both the model 1 and model 2 inputs. As outputs, it takes the merged layer. The only thing left to do is compile it and train.10. Let's practice!
It's now time to put what you've learned to work in some exercises.Create Your Free Account
or
By continuing, you accept our Terms of Use, our Privacy Policy and that your data is stored in the USA.