TGS Salt Challenge: my solution (top 20%)

This challenge aims to build a model that can detect large salt deposits from image of seismic scans. Detecting salt is important beacause in most cases large accumulation of oil and gas are also found underneith these salt beds. Salt is also much easier to detech that oil and gas and so if a model can be trained to detect these salt beds it will be very profitable for the Oil and Gas companies because it makes oil discovery faster.

Setup

clone the repo and install the packages with pip install -r requirements.txt

At the time being I was using a custom build of the popular Segmentation-Model-Pytorch library so you have to install that

$ wget https://github.com/jjmachan/segmentation_models.pytorch/releases/download/0.1.3.3/segmentation_models_pytorch-0.1.3-py3-none-any.whl -q
$ pip install segmentation_models_pytorch-0.1.3-py3-none-any.whl -qq

Dataset

The dataset is downloaded from the competition page. There is an EDA notebook in nbs/EDA.ipynb for you to view and understand the data better.

I’ve build the Pytorch loaders and some utility function to get started with the dataset quickly. Those can be found in the data.py and utils.py file.

I’ve used Albumentation library for augmenting the dataset but these can be easily switched out. Check the blog for more details about the performance boost augmentation gave.

Training

The trainer.py is a script that will run the trainer and save the trained model into the saved_models/ directory. You can modify hyperparameters in the script directly to try stuff out.

In addition to that I’ve also added 2 notebooks that I used for building this repo.

1. nbs/Baseline.ipynb - is the code to create a Baseline model that gives fairly good accuracy. It’s a useful starting point if you want to get started on your own.

2. nbs/TGS Training.ipynb - is similar to the trainer.py but for rapid development and prefer notebooks for that like me.

Training is fairly fast and takes at max 1 hr.

Inference

To run the inference with the trained model simple call the infer.py script or use the functions provided. This is generate the submission file in the required formate. To run the evaluation head over to the kaggle website and submit to competition.

I’ve implemented my own version of Test Time Augmentation(TTA) to give an additional boost to the model’s performance.