Loading...

Building Neural Networks with TensorFlow: A Beginner's Guide | CodeSignal Learn

Creating a Digits Recognition Model using TensorFlow

Let's start by defining a neural network using TensorFlow. We'll utilize TensorFlow's Keras Sequential API to create a Sequential model. This model allows for the easy construction of a neural network, where each layer connects to precisely one input tensor and one output tensor.

Consider the following:

tf.keras.models.Sequential(): This initiates our network model as a Sequential model.
tf.keras.layers.Dense(): This creates a densely connected layer, meaning all neurons in a dense layer connect to all outputs in the previous layer.
64: This number represents the number of neurons in the layer.
activation='relu': relu stands for "Rectified Linear Activation." It's a simple function that allows positive values to pass through, while it turns negative values into zero.
input_shape=(X_train.shape[1],): This specifies the shape of the input that the model will receive. This must be specified for the first layer in your Sequential model.
tf.keras.layers.Dense(10, activation='softmax'): Our final layer is designed with 10 neurons, one each for interpreting digits 0-9. Here, we use the 'softmax' activation function to predict the digit's probability distribution.

Given how large this network architecture is, it's hard to visualize it, but worth a try. In the image below you can see the 3 layers that are 64 neuron wide (including the input layer) and the output layer with 10 neurons. Because these are dense layers, we have 64*64 = 4096 edges connecting them so each individual connection is hard to see.

Let's further examine our model structure using the summary() method. This provides a quick overview of our model's architecture, along with a summary of the layers, output shapes, and the number of parameters.

output

The model.summary() method provides detailed model information, notably the param # in the output. It reflects the total number of trainable parameters in the model. You might notice we said there are 4096 edges between the neurons but here we see 4160 trainable parameters between the deep layers, where is the difference coming from? The difference is coming from bias term associated with each neuron. More about weights and biases in the next lesson!