My fluffiness recognition model is pretty inaccurate -_- Any quick improvements I could make?

I just finished the intro chapter in the free book and, in the fastai spirit, I decided to build my first model that recognises whether an image has something fluffy in it (yes, you read that right!). However, I can’t get the model the yield an error rate of less than 20%.

Things I’ve tried:

  • Increasing the number of images fed into it (currently 283).
  • Tweaking the number of epochs.
  • Changing the number of layers in the resnet model architecture I’m using.

Are there any other quick and easy changes I could make?

So it turns out I wasn’t naming the images correctly, which meant that the model couldn’t distinguish well between fluffy and not-fluffy images. The ResNet model then improved to a whopping 94%, and I even manage to get it to 98% by switching to EfficientNetV2. Here’s the link to the model: Google Colab