Code for the Kaggle Understanding the Planet from Space Competition.
Switch branches/tags
Nothing to show
Clone or download
Fetching latest commit…
Cannot retrieve the latest commit at this time.
Permalink
Type Name Latest commit message Commit time
Failed to load latest commit information.
checkpoints
ensembles
notebooks
planet
.gitignore
README.md
requirements.txt

README.md

kaggle-planet

Code for the Kaggle Understanding the Planet from Space Competition.

Overview

  • This was a team effort with Grant Bruer.
  • The final solution ranked 103rd of ~970 teams (top 11 percent).
  • The final solution was a small ensemble of ResNets and DenseNets using test-time augmentation and some trickery to weight the ensemble votes for each image. This was not terribly different from the winning solution, though the winner did use a neat de-hazing trick which was described on the Kaggle discussions. In the end there was not much gain in using an ensemble. IIRC, you could get 91% accuracy with a simple 8-10 layer conv-net. This is referred to as "Goddard" in our repository.
  • The repository is fairly messy; all of the models are convolutional networks implemented in Keras.
  • Grant and I made a Google Doc outlining some of the things we learned