alistairwgillespie / Projects

Repo with all Project types including: "Stanford Cars, road to 90%+"

Projects

Repo with my current projects

<- Launch Binder or share the Binder link

Stanford Cars

• Fine-grained image classification of different car models (Stanford-Cars dataset)
• The goal of this series of notebooks is to achieve 90%+ accuracy using fastai
• My current best is 89.4% accuracy

Feature Testing

• RGB Transforms: Tested whether different types of RGB transforms could improve validatoin accuracy for image classification. Randomly zeroing out 1 channel had the best improvement. Prompted after discussion on the fastai formum here: https://forums.fast.ai/t/rgb-transformations-for-data-augmentation/36876/8
• Achieved 1% increase in accuracy over baseline on Stanford-Cars dataset

Resources

Hows to customise the Fastai DataBlock: https://medium.com/@wgilliam/finding-data-block-nirvana-a-journey-through-the-fastai-data-block-api-c38210537fe4