Open AI Caribean Challenge: Mapping Disaster Risk from Aerial Imagery
Competition website: https://www.drivendata.org/competitions/58/disaster-response-roof-type/page/142/
What's been tried:
- Pretrained networks (ImageNet) as feature extractors with various classifiers. Didn't work: extracted feature vectors all clumped together.
- Retraining pretrained networks (ImageNet). Using entire network. AlexNet good, ResNet152 better.
Running scripts
Execute from root directory (i.e. add root directory to PYTHONPATH).
General process:
- Use src/training/grid_search.py to train models and find hyper-parameters.
- Evaluate using src/evaluation/evaluate.py. Update models/results.md.
- If keen to submit, run src/evaluation/submit.py.
Intended Project Structure
Based on https://drivendata.github.io/cookiecutter-data-science/#example
Root
|── README.md <- The top-level README for developers using this project.
├── requirements.txt <- The requirements file for reproducing the analysis environment.
|
├── data
│ ├── interim <- Intermediate data that has been transformed.
│ ├── processed <- The final, canonical data sets for modeling.
│ └── raw <- The original, immutable data dump.
│
├── docs <- Project documentation.
│
├── models <- Trained and serialized models, model predictions, or model summaries.
| └── features <- Features extracted from models.
│
├── notebooks <- Jupyter notebooks for exploration and communication.
│
├── references <- Data dictionaries, manuals, and all other explanatory materials.
│
├── reports <- Generated analysis as HTML, PDF, LaTeX, etc.
│ └── figures <- Generated graphics and figures to be used in reporting.
│
├── src <- Source code for use in this project.
| ├── setup.py <- Python installation script.
│ │
│ ├── data <- Scripts to download and extract data.
│ │
│ ├── evaluation <- Scripts for evaluating and creating submission from models.
| |
│ ├── features <- Scripts to turn raw data into features for modeling.
│ │
│ ├── models <- Scripts to train models and then use trained models to make predictions.
│ │
│ └── visualization <- Scripts to create exploratory and results oriented visualizations.
|
└── submissions <- Submission files for competition upload