Neural Beacon Placement

This code accompanies the paper Jointly Optimizing Placement and Inference for Beacon-based Localization.

Arxiv Link: https://arxiv.org/abs/1703.08612

Dependencies: Numpy, Tensorflow, and Matplotlib (for visualizations)

Important Files:

src/experiments - This directory contains files defining the parameters for each experiment. Newly created experiment files should be placed here.

src/config.py - This file defines the paths used for saving data, model weights, and results.

Evaluate a pretrained model:

We provide 6 pretrained models you can use to reproduce our results. Download the models here.

To evaluate a model, run the following commands:

unzip path_to_pretrained_models.zip
cd src
python gen_test_data.py ../maps/map1.txt #~200MB for each map
python eval_model.py anneal_map1
python gen_viz.py anneal_map1

Replace "anneal_map1" with another experiment name to evaluate other models.

Since the propagation model is noisy, your numbers may differ slightly from ours.

Train a new model:

To train a new model, create a new experiment file in the src/experiments directory. Then, run the following commands:

cd src
python gen_train_data.py ../maps/map1.txt #~3GB for each map
python run.py exp_name #Replace "exp_name" with the name of your experiment

Generate a new map:

To use a new map, convert the map to a .svg file. Then, run:

cd maps
python svg2txt.py path_to_svg 25 25 > path_to_map.txt # creates an evenly spaced grid of 25 x 25 beacon locations

To use the map, first generate train and test data:

python gen_train_data.py path_to_map.txt
python gen_test_data.py path_to_map.txt

Then, in an experiment file, set:

MAPFILE = "path_to_map.txt"