Skip to content

Latest commit

 

History

History
115 lines (73 loc) · 4.82 KB

repositories_dependencies_and_pipeline.md

File metadata and controls

115 lines (73 loc) · 4.82 KB
Raw

Repositories, dependencies, and pipeline management

This document describes the general structure and process for the full police-eis project.

What are the repos?

There are a number of repositories that are and have been used in the course of developing the police early intervention system. This document is designed to describe the purpose and content of each. There is detailed documentation about each of the repositories used.

How do I get them?

For the police-eis repository, simply run:

git clone --recursive https://github.com/dssg/police-eis.git

For the police-eis-private repository, change your directory to be in the root level of the police-eis repository, and then run:

git clone --recursive https://github.com/dssg/police-eis-private.git

The --recursive flag is important because it will make sure to clone any submodules that are included in the repositories (currently this is only pg_tools).

Dependencies

1. Drake

Only required for loading the Nashville data (see below).

2. Python 3.4+

For package dependencies, see requirements.txt. NB: luigi will not work with Anaconda because its builds are out of date and sometimes the Python path is problematic.

3. Luigi

pg_tools

pg_tools needs to exist in the repositories when you run any of the luigi population scripts. If you cloned the repositories with the --recursive flag, you should have them already. If you don't already have them, or you're getting errors that luigi cannot find pg_tools, you can try recloning pg_tools with the following commands:

cd [path to police-eis repo]\schemas\pg_tools
git submodule init
git submodule update

and for the police-eis-private repo:

cd [path to police-eis repo]\police-eis-private\schemas\pg_tools
git submodule init
git submodule update

Pipeline

Raw to ETL

We use Drake to transfer the raw data from the department to the ETL schema (only the MNPD data uses this system). To run this process use the following command in the [path to police-eis repo]\police-eis-private\schemas\etl\ directory:

drake -w Drakefile_[department] -s ~/.s3cfg

ETL to staging

We use luigi to move data from the ETL schema to the staging schema. Full documentation for this process (including repopulation) is available. There is additional detailed documentation on luigi and our setup.

Much, but not all, of the ETL to staging process has been automated with a bash script called 'run_luigi.sh' in the police-eis/schemas/ directory.

If the user would like to run this largely automated script on a staging-development schema, the schema can be specified in the first line of this file:

export schema=staging_dev

Change staging_dev to an alternate name if desired, and from the terminal run: bash run_luigi.sh.

Staging to features

To generate features, set the configuration of features (and labels) in a configuration file that has the same form as example_officer_config.yaml. And then from the police-eis directory, run:

python -m eis.run [configuration file name] --buildfeatures

Features to results

Once the features have been built, you can run all of the models with:

python -m eis.run [configuration file name]

Model interrogation, plotting, and analysis

The models are stored in the results schema. Within this schema, there are a number of tables:

  1. models
    contains all of the model information and configuration details
  2. evaluations
    contains a number of metrics for each model. Each row is a metric for one model. The model_id is a foreign key to model_id in the models table.
  3. predictions
    contains the predictions for each unit (officer or dispatch) that was predicted for each model. The model_id is a foreign key to model_id in the models table.
  4. data
    contains model pickle files that can be read back into python if necessary.

Reading model configuration information

The entire configuration file for every model run is stored in results.models.config as a JSON object, which can be queried directly in PostgreSQL. An example of such a query is shown here:

select * from results.models where config->>'fake_today' = '01May2014'

Reading pickles

In order to use a fit model in python (for visualization of ROC curves, model inspection, making predictions with a fitted model, etc.), simply query the database for the pickle, and then use pickle.loads()

import pickle

[sql-cursor].execute("SELECT pickle_blob FROM results.data WHERE model_id = 20")
result, = cur.fetchone()
data = pickle.loads(result)