Sign in to continue learning
Okay, so you just saw how to create a neural network that gives basic computer vision capabilities to recognize different items of clothing. Let's now work through a workbook that has all of the code to do that. You'll then go through this workbook yourself, and if you want, you can try some exercises. Let's start by importing TensorFlow. I'm going to get the fashion MNIST data using tf.keras.datasets. By calling the loadData method, I'll get training data and labels as well as test data and labels. For more details on these, check back to the previous video. The data for a particular image is a grid of values from 0 to 255 with pixel grayscale values. Using matplotlib, I can plot these as an image to make it easier to inspect. I can also print out the raw values so we can see what they look like. Here you can see the raw values for the pixel numbers from 0 to 255, and here you can see the actual image. That was for the first image in the array. Let's take a look at the image at index 42 instead, and we can see the different pixel values and the actual graphic. Our image has values from 0 to 255, but neural networks work better with normalized data. So let's change it to between 0 and 1 simply by dividing every value by 255. In Python, you can actually divide an entire array with one line of code like this. So now we design our model. As explained earlier, we define the input shape of the data. Then we define the first layer which flattens the input. The output dense layer will have 10 units to match the number of classes. The dense layer in between will try to figure out the rules between the inputs and the outputs. Now we compile the model defining the loss function and the optimizer, and the goal of these is, as before, to make a guess as to what the relationship is between the input data and the output data, measure how well or how badly it did using the loss function, use the optimizer to generate a new guess, and repeat. We can then try to fit the training images to the training labels. We'll just do it for five epochs to be quick. We spend about 25 seconds training it over five epochs, and we end up with a loss of about 0.29. That means it's pretty accurate in guessing the relationship between the images and their labels. That's not great, but considering it was done in just 25 seconds with a very basic neural network, it's not bad either. But a better measure of performance can be seen by trying the test data. These are images that the network has not yet seen. You would expect performance to be worse, but if it's much worse, you have a problem. As you can see, it's about 0.33 loss, meaning it's a little bit less accurate on the test set. It's not great either, but we know we're doing something right. Your job now is to go through the workbook, try the exercises, and see by tweaking the parameters on the neural network or changing the epochs if there's a way for you to get it above 0.71 loss accuracy on training data and 0.66 accuracy on test data. Give it a try for yourself.