Thread by @matteofigus

Days ago AWS announced #Rekognition Custom Labels: a new feature that allows you to detect objects inside images with little Machine Learning expertise. I decided to test it for solving one of the most important problems for me: classify proper #pizza. A #pineapple thread 1/13

I am not new to ML, but being a SW dev for most of my career, I always found hard to:

1. keep up with all the new frameworks and algorithms
2. find the best algorithm for my data & problem
3. understand and configure all the parameters needed for starting the training

2/13

With the new Rekognition Console, you start by having some images, then you can manually label and draw bounding boxes. I started by collecting ~50 pizza images and saving them to a S3 bucket. Obviously I included real pizzas and fake pizzas. What's a fake pizza?

3/13

If you are not a pizza expert it may be a hard concept to grasp. I'll explain it with an example. Can you order a black coffee with milk in a bar? No. Because it doesn't exist. If a black coffee has milk, by definition it ceases to be a black coffee. Simple.

4/13

Now let's get back to pizza. For pizza the same applies when #pineapple and #corn are involved. I decided to create some labels in the Console in a new dataset: "Proper pizza", "Fake Pizza", "Corn", "Pineapple". Then I spent ~10m to manually label all the images.

5/13

When I was done with labelling I started the training by literally clicking a button and I went back a couple of hours later and it was ready. Then all you need to do is start the model and use an API to test your model.

6/13

To make testing simple, I worked on a code sample that helps with that: github.com/aws-samples/am… - it allows to see all the models and test the ones running. You can deploy it with one click.

7/13

Here is my first test. Pineapple label detected, good. No pizza detected, ok I agree with that, well done machine, we can be best friends.

8/13

Here is a good looking pizza. Oh, look at this beauty. Proper pizza detected: yay. Success!

9/13

I tried to test another 20/30 images and found that the proper vs fake pizza worked fine: 100% accuracy with my manual tests!

10/13

In conclusion, I'm pretty happy with my test. With my limited ML knowledge I was able to elevate my pineapple pizza trolling to new levels. I also learned a couple of things about accuracy and how it relates to labelling, and I'll keep playing with the tools.

11/13

If you play with Rekognition console, remember you spend money for all the models in the "RUNNING" state (so remember to stop them when done to avoid wasting money). Also, check this little tool for having a simple UI to manage and test your models: github.com/aws-samples/am…

12/13

Final little disclaimer: opinions are my own. Especially those about pizza.

13/13