salayatana66 / vw_utilities

Utilities for Machine Learning with Vowpal Wabbit

Utilities for Machine Learning in Vowpal Wabbit

In this repo I plan to add utilities to work with Vowpal Wabbit.

• vwUniqueFeasExtractor.py is a program that summarizes features and interactions in the training set to produce a new test set on which hash inversion is faster. The reason to write this is that utl/vw-varinfo in Vowpal Wabbit does not ingest UNIX pipes. Working as a DS in a production environment where hash inversion is required on large datasets I have used it to ingest UNIX pipes (say resultin from hive queries) while dumping the results into other pipes.

  Example:

  chmod +x vwUniqueFeasExtractor.py

  ./vwUniqueFeasExtractor.py --interactions 'a*b,a*b*c' <inputFileOrPipe > outputFileOrPipe

• tensorflow_wrapper is a little python package to wrap a model into a tensor flow one, e.g. for serving a large model without need to explicitly encode interactions. Currently the weights must be human readable, in a file of the form

  Weight Name   Hash    Weight Value
  

  with the header removed.

  • Step 1: Import the wrapper package, e.g.

  from tensorflow_wrapper.Wrapper import FeaType as fT, VowpalWabbitWrapper as WP
  

  • Step 2: Create a nested dictionary of namespace, features and category type, e.g.

  feaDict = {"a" : { "a1" : fT.numerical, "a2" : fT.categorical},
  	  "b" : {"b1" : fT.categorical} }
  
  

  • Step 3: instantiate the wrapper, e.g.:

  myWrapper = WP(feaDict, interactionsString="a*b,a*a")
  

  • Step4: load the weights from a file, and wrap

  myWrapper.loadWeightsFromFile("file.tsv")
  myTensors = myWrapper.wrapModel()
  

  Then input is available in

  myTensors["Input"]
  

  as a dictionary on the feature name. And the response of the model is available as

  myTensors["response"]