kaggle-cassava
Team blue (formerly team dumpsterfire) submission for the Cassava Leaf Disease Classification competiton on kaggle, including our model, training methods and submission methods. This repository is meant to serve as an addition to the report for our period 3 project supervised by Chris Pawley at MSP. As of now, we have achieved an accuracy of 89.0 on the kaggle testing data, placing us at 1661/3105 on the leaderboards. All the code found here is free to use under the specified liscense, and is documented. Our final notebooks can be found under the project folder. The don't do this folder serves to show off our worst model.
In order to use this code in a kaggle notebook the appropriate datasets must be imported: https://www.kaggle.com/dimitreoliveira/efficientnet-git - To use the efn model
https://www.kaggle.com/dimitreoliveira/kerasapplications - pretrained models and weights
https://www.kaggle.com/itsuki9180/cassava-recreate-stratificated-tfrecords - stratified dataset
https://www.kaggle.com/dimitreoliveira/cassava-leaf-disease-tfrecords-512x512 - cleaned tfrecords
https://www.kaggle.com/dimitreoliveira/cassava-leaf-disease-tfrecords-center-512x512 - cleaned tfrecords, centre cropped