During my winter break, I took on the Cassava Leaf Detection competition on Kaggle, which I thought would be a quick project that would keep me occupied. However, I found out I was greatly mistaken. What I thought would be a quick few week project ended up taking me almost three months. I started the project at the start of my winter break, December 23, and my last submission was made on the day that the competition ended, February 11.

Initially, I had hoped I could make a simple model based off the ResNet architecture because I have had success with that before. These hopes were crushed when I found I could get no more than 70% accuracy. I then decided I needed a more complex model and decided to use transfer learning. I decided to use the pre-trained ResNet50 model. To my surprise, This performed worse than my initial model. I then tried to fine tune my initial model and was able to hit 75% before overfitting. I then tried both the pre-trained Inception and MobileNet architectures. The Inception model was only able to hit 70% without any fine-tuning and I quickly abandoned it. The MobileNet architecture was able to hit 80% with a lot of fine-tuning but I decided to drop it because I did not think I could get any more out of it. So, late in to the competition, I was scrambling to try and find an architecture that could work.

I finally stumbled onto the pre-trained EfficientNet model. My first test I was able to hit 83% before overfitting. The only problem was, for a reason I have yet to discern, the model would sometimes decide not to learn and a training session would be wasted, also wasting my precious, limited GPU hours. It was the final week of the competition and, furiously rushing, I was able to tune the model to get 86%. With my GPU hours gone and the competition coming to a close I was able to submit my final model with a score of 0.8611. I got in 2854th out of 3900 teams, the top team getting a score of 0.9132.

A visualization of the model

This was my first Kaggle competition and I was able to learn a lot. If I were able to redo this, I would not have spent as much time fine-tuning the MobileNet model as most of it didn’t lead anywhere. Overall, I am happy with this project and I think I learned quite a bit.

I meant to get this post out during February or March but I was still revamping the website and I didn’t finish until March 19th which was the end of my Spring Break. With the end of the school year I wasn’t able to work on any personal projects and I was only able to work on this post at the start of summer. I also have a recap of the extended summer, due to quarantine, of 2020 that I need to post and I hope to post that soon.