I love working with AI to find new solutions. I thought about a field that is yet to adopt AI and thought of apps such as myFitnessPal or Calorie Tracker.
What it does
PIK is a modern day food app that solves the tedious process of manually inputting your meals. The user will only need to take a picture and an AI model will predict what food is present on that picture and how much calories it has.
How I built it
I have managed to only build the initial AI model for food classification. Using pytorch, I made use of transfer learning to accelerate the development process and save myself the trouble of figuring out why my model isn't training.
Challenges I ran into
At first, I didn't achieve very good training results, but I managed to solve that by trying out different optimizers, loss criterions, and image preprocessing methods. I also spent quite a lot of time trying to figure out how to use AWS SageMaker to train the model before realising that it might be better to train on Colab on a smaller dataset. I am yet to make use of EC2 for the full training and deployment of the model.
Accomplishments that I'm proud of
I am proud that I managed to actually train the model, even though it is only the first step to developing the app itself.
What I learned
Despite not making use of it, I think I learned way more about cloud computing than I did previously, especially because I was able to see just how much I have yet to learn.
What's next for PIK
The next step for PIK is training the full model, preferably on AWS so that I can deploy the model. To create the calorie tracking functionality, I need to develop a backend app which would make use of a food api containing all the necessary calorie and nutrition information. Lastly, I need to develop the frontend application itself.