Skip to content

Generates useful feature embeddings for geospatial locations.

Notifications You must be signed in to change notification settings

JakeForsey/geobacter

Repository files navigation

Overview

Geobacter generates useful location embeddings on demand, it is an implementation of the Loc2Vec blog post from sentiance

A resnet is trained to embed renderings of geolocations using the triplet loss. Samples are generated based on the principle that:

"Everything is related to everything else, but near things are more related than distant things"

Anchor Positive Negative

Setup

Initialise the open street map tile volumes and server

docker volume create openstreetmap-data
docker volume create openstreetmap-rendered-tiles

docker run \
    -e THREADS=12 \
    -v $PWD/data/osm/luxembourg-latest.osm.pbf:/data.osm.pbf \
    -v openstreetmap-data:/var/lib/postgresql/12/main \
    overv/openstreetmap-tile-server \
    import
export PYTHONPATH=$PYTHONPATH:$PWD/geobacter

Create a python environment (for training)

pipenv install --dev
pipenv shell

Create a python environment (for inference)

pipenv install
pipenv shell

Start the open street map tile server

docker-compose up

Train

Initialise some training and testing samples (which also caches tiles)

python bin/generate_samples.py --sample-count 100000 --buffer 100 --distance 500 --seed 1 --path data/extents/train_100000.json
python bin/generate_samples.py --sample-count 10000 --buffer 100 --distance 500 --seed 2 --path data/extents/test_10000.json
python -m geobacter.train

Run

(optional) Check that the open street map tile server is up

curl localhost:8080/tile/16/33879/22296.png --output test.png

Start the python service

export GEOBACTER_TOKEN=<token>
gunicorn -b 0.0.0.0:8000 --workers 4 --timeout 10 geobacter.inference.api:app

(optional) Get the embedding for Notre-Dame

curl "localhost:8000/embeddings?lat=49.609598&lon=6.131606&token=<token>" | jq
{
  "embeddings": [
    0.12629294395446777,
    0.5683436393737793,
    0.9822958111763,
    0.38620898127555847,
    -1.2079272270202637,
    0.16978177428245544,
    -0.3008042275905609,
    0.06522990763187408,
    0.5405853390693665,
    -0.8018991947174072,
    0.42124632000923157,
    0.6691603064537048,
    -0.40959250926971436,
    -0.18567749857902527,
    -0.017753595486283302,
    0.3173545002937317
  ],
  "checkpoint": "checkpoints/ResNetTriplet-OsmTileDataset-e393fd34-aa3c-4743-b270-e7f0d895b0a8_embedding_41450.pth",
  "lon": 6.131606,
  "lat": 49.609598,
  "image_url": "image?lon=6.131606,lat=49.609598,token=<token>"
}

Results

Semantically similar locations are embedded together

The embedding space can be interpolated

Similar locations can be queried

Examples

Use the api to characterise a pre-created route.

examples/api.py

Use a checkpoint to characterise a large number of samples.

examples/checkpoint.py

About

Generates useful feature embeddings for geospatial locations.

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages