Small web app deploying a CNN built with the fast.ai library
CNN built using the Fast AI library. Model specifications:
- Overview: CNN used to classify oval sports balls into Rugby Balls, Australian Rules Footballs or American Footballs
- Architecture: ResNet34 (with optimised hyperparameters)
- Data: Training set conisted of 575 images downloaded from Google Images (143 in the validation set). Datasets excluded images that were not photographs of oval sports balls (e.g. hand drawn images, clip-art, GIFs)
- Training: ~10 minutes on a GPU hosted on Colab; 94.4% accuracy
- Supporting materials: Largely follows lesson 2 of the Fast AI Practical Deep Learning for Coders (v3) course
- Further reading: More on prolate spheroids here
- Note: Rugby Union and Rugby League balls treated as one class
Sample images: Wikipedia