/schneider_zero_deforestation

Big and diverse ensemble scoring 0.75 on validation set

Primary LanguageJupyter Notebook

Schneider's Zero Deforestation Hackaton

Cambodia deforestation

Predict deforestation on satellite images

https://nuwe.io/dev/challenges/data-science-se-european

Docs

Enjoy the documentation of the project created with MkDocs at https://ironbar.github.io/schneider_zero_deforestation/

Solution

As required by the challenge the solution has 3 files:

  • main.sh
  • predictions.json
  • presentation.pdf

Solution summary

An ensemble of the best 13 fine-tuned models comprising of:

  • OpenClip models pretrained on LAION dataset (ViT-B-16-plus-240, ViT-B-32, ViT-L-14, ViT-H-14, ViT-g-14, ViT-B-16)
  • Keras models pretrained on Imagenet (ResNet50, ResNet50V2, MobileNetV2)
  • Achieves a validation f1 score of 0.754

Code structure

 |_ docs: documents made during the challenge according to CRISP-DM methodology
 |_ models: predictions and models trained for the challenge
 |_ notebooks: jupyter notebooks made during the challenge. They start by number for easier sorting.
 |_ rules: the official rules of the challenge
 |_ scripts: scripts made during the challenge for training, data processing...