Programming PyTorch for Deep Learning

Chapter 3. Convolutional Neural Networks

After experimenting with the fully connected neural networks in Chapter 2, you probably noticed a few things. If you attempted to add more layers or vastly increase the number of parameters, you almost certainly ran out of memory on your GPU. In addition, it took a while to train to anything resembling somewhat decent accuracy, and even that wasn’t much to shout about, especially considering the hype surrounding deep learning. What’s going on?

It’s true that a fully connected or (feed-forward) network can function as a universal approximator, but the theory doesn’t say how long it’ll take you to train it to become that approximation to the function you’re really after. But we can do better, especially with images. In this chapter, you’ll learn about convolutional neural networks (CNNs) and how they form the backbone of the most accurate image classifiers around today (we take a look at a couple of them in some detail along the way). We build up a new convolutional-based architecture for our fish versus cat application and show that it is quicker to train and more accurate than what we were doing in the previous chapter. Let’s get started!

Our First Convolutional Model

This time around, I’m going to share the final model architecture first, and then discuss all the new pieces. And as I mentioned in Chapter 2, the training method we created is independent of the model, so you can go ahead and test this model out first and then come back ...