The goal of this competition was to identify and segment salt deposits in
underground seismic images. For an overview of the project and my approach, see
overview.ipynb. The official Kaggle page can be found
here.
I placed 384th out of 3,234 teams, which is in the top 12th percentile. According to Kaggle, there were nearly 4,000 participants, making it one of their biggest competitions of the year. My Kaggle profile can be found here.
The following files contain all of the code used to make the predictions:
-
run.pycan be run from the command line to load the data, train the models, and make predictions.config.jsoncan be passed as a command-line argument if you want to change the parameters. -
src/data.pycontains the functions necessary to construct the train and test datasets from the images, masks, and other raw data. -
src/model.pyfully implements the deep residual network I built for this competition in Tensorflow. -
src/utilities.pycontains several functions used by the model. This includes the specialized loss functions and a custom accuracy metric which matches the competition's scoring mechanism.
The directories data/ and output/ are omitted from this repository. To
download the data and create these directories automatically, run setup.sh
from the command line.
The Jupyter notebook overview.ipynb gives a brief overview of the project and
the image data.
Lastly, papers/ contains several research papers I used as references for my
techniques and model architecture. All of these are freely available on the
internet.
This project is designed for Python 3.6+. The simplest way to get the required packages is with an Anaconda distribution. In addition, you must install the Kaggle API to obtain the data.
Then, clone this repository, run setup.sh to obtain the data, and run
run.py to generate predictions.
Lastly, you can also adjust the training parameters in the configuration file
and run python run.py config.json to use those settings.