Notebook used for the Kaggle competition PetFinder.my - Pawpularity Contest
This notebook especillay showcase how transfer learning can be applied based on the EfficientNet model.
It also showcases how we can create a custom generator from Keras's Sequence class so that we can avoid loading all the images in RAM which would cause out-of-memory errors.
We create a model based on two input, the first one being the RGB vector from the given pictures and the second one being a vector representing some metadata that have been manually labelled by the organizer of the competition.
Our best score with the presented approach is a RMSE of 18.81 that allowed us to rank ~Top50% on the public leaderboard (temporary standings as the competition is not over as this notebook is being published).