Repository contains code to solve Kaggle problem Planet: Understanding the Amazon from Space. This solution won 3rd place in competition.

Requirements:

Python >= 3.4, Keras 1.2.1, Theano 0.9.2, Tensorflow, XGBoost 0.6

How to run:

You need to execute set of scripts one by one:

• python a11_find_neighbours.py
• python a30_create_keras_models.py
• python a30_create_keras_models_land.py
• python a30_create_keras_models_weather.py
• python a30_create_keras_models_single_class.py
• python a31_create_cnn_features_basic.py
• python a31_create_cnn_features_land.py
• python a31_create_cnn_features_weather.py
• python a32_create_cnn_features_single_class.py
• python a32_find_neighbours_features.py
• python a42_gbm_blender.py
• python a42_keras_blender.py
• python a50_ensemble_from_cache_v1.py

Notes:

1. Recreating all CNN models from scratch on single GPU will require a lot of time (around a month). It can be parallelized using separate GPU on different CNN models. Final models weights size ~50 GB. Msg me if you need these weights.
2. Creating neighbours features requires around a day to complete.
3. Due to high parallelization, CNN models trained on GPU can slightly differ even in case it was trained on the same code.
4. A little bit details about solution available on Kaggle forum

Directory structure:

• -- input - input data as it was given on Kaggle
• -- Kaggle-Planet-Understanding-the-Amazon-from-Space - all the Python code (this repo)
• -- models - all generated models from neural nets will be in this folder.
• -- weights - files with weights for pretrained models. Link: Download
• -- modified_data - some intermediate files for neighbour analysis
• -- features - all raw features generated by neural nets will be stored in this folder. We already have them calculated. Link: Download
• -- cache - this folder will contain arrays with predictions from XGBoost and Keras blenders
• -- subm - final predictions (in format of submit file for Kaggle)

Dataflow