I did lesson 1 and I’m currently just practicing.
First, I worked on Image Classification to distinguish coyotes from foxes. There are 25 images of each animal. The dataset is here and the notebook is here. This classification was pretty much 100% accurate, so I was wondering whether I was overfitting because I’m not completely sure how to know when I am.
I decided to try something bigger, an image classification of car body styles. There are 250 images for each of the following body styles: Convertible, Coupé, Hatchback, Minivan, Pickup, Sedan, SUV and Truck. The dataset is here and the notebook is here. In this one however, with resnet34, I got 62% accuracy and then 63% after unfreezing and fine-tuning. With resent50 I got 65% accuracy and then 70% after unfreezing and fine-tuning. I was amazed I even got 70% but I was pretty much just using the lesson 1 notebook as a template.
I did work on improving the dataset such as by checking to make sure I got good images and making sure each picture had only one car.
Is there anything I can do to try to improve the accuracy for the second project before I start lesson 2?