To wrap up the course, each student creates a capstone project; something to demonstrate the knowledge gained over the past twelve weeks. We don’t have many guidelines and are encouraged to be creative with our problem statement.

Recently, I adopted a puppy and was told its mother was a Boston Terrier, but that’s all they know. I have since been to two separate vets who guess that my puppy, Boomer, might also be parts Boxer and/or Beagle. I have google image searched both Boston-Boxer and Boston-Beagle mixes and can see traces of him in both. However, many people seem pretty convinced of either Beagle or Boxer. My vet told me that for a cheap $100, I could find out the DNA, but since it’s a dog and spending $100 on a DNA test won’t provide much use other than being able to provide a concrete answer to the oft questioned ‘What kind of dog is he?,’ I figured I could end all suspicion through the creation of a neural network.

I initially had great ambitions to classify nearly 120 breeds, using an Image-net database of over 100,000 images. I quickly found that a GPU can’t handle that, and had to narrow down my data to just a few breeds. It seemed pretty obvious to me that I should just take Boston Bull terriers, Beagles, and Boxers, each with a little over 1,200 images.

As this was my first endeavor into a neural network, I did a great deal of research and spent a lot of time processing my images. Image-net provides bounding boxes for some of their synsets of images, and luckily each of my three chosen breeds had them. I hadn’t originially bothered with them, because when I was using all breeds, I found that only some had bounding boxes, and not all. So after a lot of struggling, I realized that I should double check for my three breeds.

Image-Net is a commonly used image database, so I was hoping that there would be tons of tools shared to easily download images and bounding boxes, but I didn’t find them until after I figured it out on my own. I ended up trying the utilties from this repo:. It’s an excellent resource, and will even crop the images according to their bounding boxes.

Unfortunately, not all of the photos per sysnset have bounding boxes, which I didn’t realize beforehand. The process from ImageNet_Utils had a separate folder of original images and one with the cropped images, and due to the difference in naming conventions, it would’ve taken more time that I had available to determine which images didn’t have bounding boxes and should just be included anyways. Because of this, the number of images I had for this projects was greatly reduced.

Luckily, there are plenty of resources online detailing how to deal with small sets of images–perform image augmentation. Essentially, more images are created from the existing ones by flipping them, rotation, shifting, etc. so that the model can have more of a variety of images to train on.

As I have never built a nueral network, the modeling was a bit tough for me. I found several examples of different layers to add or pretrained models, but I settled on my final model because the accuracy scores from training were nearly identical to what the test score evaluated to, ~67%. My other models were quite overfit and although training scores were high 90%s, they dropped nearly 30% with the test data.

To conclude my project, I ran a set of images of Boomer through my model to assess the percentages of each breed that were predicted. Turns out, it’s pretty inconclusive. I guess if I really want to know, I should just fork over the $100.

Check out the whole capstone piece here.