Skip to content
Command line interface for the ml-enabler
Branch: master
Clone or download
Fetching latest commit…
Cannot retrieve the latest commit at this time.
Type Name Latest commit message Commit time
Failed to load latest commit information.
ml_enabler validate aggregate zoom cannot be higher than source, fixes #34 Jun 26, 2019
.flake8 more linting, add flake8 config (allow 160 chars per line) Jun 18, 2019
.gitignore test looking glass aggregator Jun 17, 2019


Command line utilities for working with ML models and ml-enabler-api


ml-enabler-cli is a library to interact with ML Models, run prediction jobs, and submit results to the ml-enabler-api. The CLI also has a utility to query OSM (only buildings now), and aggregate data at a defined tile level. See ml-enabler-cli for more information.

At the core of the library there are Predictors and Aggregators to work with models, and a command to upload predictions.

  • Predictors Predictors interact with a model hosted at a particular endpoint. At the moment, we've integrated Development Seed's looking-glass model and a building API that returns geometries for a given bbox. To see all the available predictors, see here.

  • Aggregators Aggregators take predictions from a model and then aggregate data at desired zoom level or tile size. The output is in the format acceptable by the ml-enabler-api. Aggregators can optionally query Overpass for OSM data and add that to the predictions. See the looking-glass aggregator for example.



  • Git clone this repo
  • cd ml-enabler-cli
  • Create a virtualenv (recommended) python3 -m venv env
  • pip install -e .
  • Install osmtogeojson (this is used for matching with OSM data in the aggregator) npm install -g osmtogeojson
  • ml-enabler-cli --version to check if the installation was successful
  • ml-enabler-cli --help to see commands and options

Looking Glass

To use looking-glass, first ensure looking-glass is hosted either on your computer or on a GPU cloud instance. Instructions to this are here.

Fetch predictions

To fetch predictions for a bbox, use:

ml-enabler fetch_predictions --name looking_glass \ 
 --bbox "-77.14, 38.82, -76.92, 38.95" \
 --endpoint \
 --tile-url{z}/{x}/{y}.jpg?access_token={token}' \
 --token abcd \
 --zoom 18 \
 --outfile /tmp/looking_glass_output.json \
 --errfile /tmp/looking_glass_errors.json

The predictor options are:

  • bbox - bbox required to fetch predictions for.
  • endpoint - URL where the model is hosted. The model should be following the TFServing schema
  • zoom - zoom level to be predicted at. This is usually determined by the model
  • tile-url - tile url to pass to the model. This is usually determined by the model, and will be satellite imagery. For models/apis that don't need tile-url it's safe to ignore
  • token - access token for the fetching tiles from the tile-url
  • lg-weight - custom weight parameter. default is auto, and select the weight defined by the model
  • concurrency - number of concurrent http requests to make (to the model as well as the tile server)
  • outfile - output file to store the predictions
  • errfile - in case certain tiles fail to predict, these are stored in the errfile for further inspection

Aggregation and fetching OSM data

To aggregate predictions, use:

ml-enabler aggregate_predictions --name looking_glass \
 --zoom 16 \
 --infile /tmp/looking_glass_output.json \
 --outfile /tmp/looking_glass_aggregated.json

The aggregator options are:

  • zoom - zoom level to aggregate the predictions to. For example, if the predictions are at zoom 18, and if you want to aggregate them to zoom 14, specify 14
  • overpass-url - optional overpass url. default is
  • infile - input file with predictions. This is the output of fetch_predictions command
  • outfile - file to store aggregated predictions

Working with the API

The ml-enabler-api acts as a gateway between models and mapping tools like Tasking Manager. The extendable API provides methods to fetch predictions at a bbox level as well as for custom polygons from a GeoJSON. The CLI allows to post results prepared from models to the API.

To upload predictions, the model should be registered first. To register the model, check the API docs.

The model_name is available in the metadata object of the prediction generated from the fetch_predictions command. ml-enabler-cli can upload predictions, using:

ml-enabler upload_predictions \
 --infile /tmp/looking_glass_aggregated.json \

Adding a new model

To fetch predictions from a new model, it is required to add a Predictor. The base class BasePredictor offers all basic options and handles fetching command defaults. Once you create custom logic as a class, add this to predictors/ Then the fetch_predictions can use the name to map the command to the new Predictor. See LookingGlassPredictor for example.

Adding a new aggregator

Similar to a Predictor, the BaseAggregator offers a basic interface to work with aggregator logic. To create a new aggregator, subclass the BaseAggregator and write custom logic, and register this in aggregators/ The aggregate_predictions can use the name to map the command to the aggregator. See LookingGlassAggregator for example.

Development Setup

  • python3 -m venv env - create a virtualenv
  • Run pip install -e .
  • Run ml-enabler --help to see a list of commands and options
You can’t perform that action at this time.