Releases: dssg/police-eis
Releases · dssg/police-eis
April 10th 2017 release
Iterate blocks (#215) * change get dataset to only return a df, and storing evaluations using model_id * test using triage * traige working with many test_dates and using metta to store matrices * change the way we call the dates for generating the features * use metrics for percentage and absolute values for evaluation * added parallel execution of feature generation * updated result schema with indexes * change metadata to datetime * change logistic regression parameter C_reg to C * add test_time_ahead * add raw incident type and levels * change group incident to complaint type to add raw type * delete grouped incident type or incident type levels from incidents table, just leaving the raw code * create functions to create labels table from config file and reads the matrices for loading train and test matrices * change example config file * add dispatch type code for original and final response and add new categories for the dispatch lookup table to match for cmpd * remove features using group incident type code * added new field interview outcome type banned * added new field interview types * bracket was missing at dispatch config * fixed bug with unclosed bracket and prepeard config * updated config and featuers for new categories and limited selection based on 1600 colum limit * changed config to be below 1600 columns * merge with new configuration for labels and create function for generating all matrices * delete events with NULL event_datetime and officer ids from labels table * changes in config file for labels * added locking function which is NFS save * add make_hashable function for metadata * removed debugging statements * changed data-types to real in results to save space and added a timeout for locking * fix hash function, DaysSinceLastCompletedAllegation and config file * fix labels table removing null values * make parallelize generate matrices * use psycopg2 for reading features and naming cursor * changed model running to train a single model first and then test across all test sets and then discard the model * add labels_config to models_group * added a delete statement before new evaluations for the same as_of_date and model_id are written to the database * add officer_id column and labels_config that passes into metta * change query for grabbing labels to exclude as of date + prediction window interval * add model_config to metadata for creating model_group_id * added rounding for storing float to avoid underflow errors in postgres * increased verbose output * add function for generating as of dates for feature generation * typo * change officer roles table adding dates and department * comment util functions and change train_metadata send to db_parameters to the config file in the db * add ucr4 code grouped matching dispatch types, add lookuptables for ucr4 grouped and codes * delete old code for feature generation, add docstrings * remove unused config file * be able to iterate through blocks leaving one block out and reading features tables directly from collate tables * add arrest crime feature and compliments block * add sworn flag * fix stored procedure for reading feature blocks * add individual importances table in production schema * iterate through blocks and add post collate feature for dispatch movement * change log info and add DispatchMovement to config * add create patrol districts table and change config parameters * stored procedure for getting complete feature description * change evaluations as of date column to start and end time * delete patrol distrincts * added outside employment * add train end time and experiment hash to results.models * add experiment hash for triage and fix evaluation start time * remove feature abstract unused classes, update config and add features descriptions