xiangyudxy / indoor-location-navigation-public Goto Github PK

This repository contains the code to generate the winning solution of the Kaggle competition on indoor location and navigation organized by Microsoft Research.

Our team name: "Track me if you can".

Authors:

• Are Haartveit
• Dmitry Gordeev
• Tom Van de Wiele

• Indoor location navigation Kaggle competition
• Summary of our winning solution
• Competition ranking

1. Clone the repository, it doesn't matter where you clone it to since the source code and data are disentangled.
2. Create a project folder on a disk with at least 150GB of free space. Create a "Data" subfolder in your project folder. This will be referred to as "your data folder" in what follows.
3. Download the raw text data from here and extract it into your data folder.
4. Download the cleaned raw data from here and extract it into the "reference_preprocessed" subfolder of your data folder.
5. Add your data folder to line 19 in src/utils.py.
6. Run main.py.

If all goes well, the pipeline should create a "final_submissions" subfolder in your data folder with two final submissions. Note that these are likely slightly different from our actual submissions due to inherent training stochasticity. When you make a late submit of these submissions to the leaderboard, you should obtain a private score around 1.5, which can be further reduced to about 1.3 after fixing the private test floor predictions (not part of this repository).

• Mode ("-m" or "--mode"). Default: 'test'. Select from ('valid', 'test').
• Suppress multipricessing ("-s"). Default: no suppression of multiprocessing.
• Fast (and bad) sensor models ("-f"). Default: no fast sensor models. Mostly useful for verifying that all dependencies are in place. Ignored when copying sensor models (next parameter).
• Copy sensor predictions ("-c"). Default: no copying of pretrained sensor predictions. Useful if you want to speed up the pipeline since training sensor models is the slowest part.

Due to the size of the data set, you need at least 32 GB RAM to be able to run the pipeline successfully.

• If you run out of memory, try running the pipeline again. It should continue where it left it in the previous run.