Fast, accurate, lightweight, multi-core ML in Python, leveraging Vowpal Wabbit
Clone or download
Fetching latest commit…
Cannot retrieve the latest commit at this time.
Failed to load latest commit information.

Vowpal Platypus

Vowpal Platypus is a lightweight Python wrapper built on Vowpal Wabbit, that uses online learning to acheive great results.

VP is...

  •, generating MovieLens predictions with just a few nanoseconds per prediction on a 40 core EC2.
  • ...accurate, acheiving AUC > 0.9 with a Titanic model that processes, trains, and predicts all in under a second on a laptop.
  • ...lightweight, with no dependencies other than Python, installing on a Macbook pro in 0.3 seconds.
  • ...multicore, scaling linearly across any number of cores, being used for hundreds of GB of data.
  • ...out-of-core, bottlenecked by CPU and IO rather than RAM.

See demo code here showing detailed implementations and benchmarks for MovieLens ALS, Criteo ad click prediction, NumerAI stock prediction, and Titanic survival.


  1. Install vowpal_wabbit. Clone and run make
  2. Clone vowpal_platypus and run sudo python install. You will also need to install Retrying, Pathos, and Dill.

If you're on Mac OSX, you will need coreutils (brew install coreutils).

(See full installation instructions if necessary.)


Predict survivorship on the Titanic using the Kaggle data:

from vowpal_platypus import run                        # The run function is the main function for running VP models.
from vowpal_platypus.models import logistic_regression # vowpal_platypus.models is where all the models are imported from.
from vowpal_platypus.evaluation import auc             # vowpal_platypus.evaluation can import a lot of evaluation functions, like AUC.
from vowpal_platypus.utils import clean                # vowpal_platypus.utils has some useful utility functions.

# VW trains on a file line by line. We need to define a function to turn each CSV line
# into an output that VW can understand.
def process_line(item):
    item = item.split(',')  # CSV is comma separated, so we unseparate it.
    features = [            # A set of features for VW to operate on.
                 'passenger_class_' + clean(item[2]),  # VP accepts individual strings as features.
                 'last_name_' + clean(item[3]),
                 {'gender': 0 if item[5] == 'male' else 1},  # Or VP can take a dict with a number.
                 {'siblings_onboard': int(item[7])},
                 {'family_members_onboard': int(item[8])},
                 {'fare': float(item[10])},
                 'embarked_' + clean(item[12])
    title = item[4].split(' ')
    if len(title):
        features.append('title_' + title[1])  # Add a title feature if they have one.
    age = item[6]
    if age.isdigit():
        features.append({'age': int(item[6])})
    return {    # VW needs to process a dict with a label and then any number of feature sets.
        'label': int(item[1] == '1'),
        'f': features   # The name 'f' for our feature set is arbitrary, but is the same as the 'ff' above that creates quadratic features.

# Train a logistic regression model on Titanic survival.
# The `run` function will automatically generate a train - test split.
run(logistic_regression(name='Titanic',    # Gives a name to the model file.
                        passes=3,          # How many online passes to do.
                        quadratic='ff',    # Generates automatic quadratic features.
                        nn=5),             # Add a neural network layer with 5 hidden units.
    'titanic/data/titanic.csv',     # File with the data (will automatically be split into random train and test)
    line_function=process_line,     # Function to process each line of the file
    evaluate_function=auc)          # Function to evaluate results

This produces a Titanic survival model with an AUC of 0.7241 (on the Kaggle holdout validation set) in 0.16sec.

Multicore Capabilities

Documentation coming soon.


Documentation coming soon.

Available Models

Nice interfaces: linear_regression, logistic_regression, als (Alternating Least Squares Collaborative Filtering) Raw interface: LDA, BFGS, Simple Nueral Nets, LRQ (nice interfaces coming soon)

Credits, Contributions, and License

The base for this repository is Vowpal Porpoise developed by Joseph Reisinger @josephreisinger, with further contributions by Austin Waters ( and Daniel Duckworth (

This software is built using an Apache 2.0 License, the license used by the original contributors.

This repository also depends and is built upon Vowpal Wabbit developed by John Langford and other contributors, released under a modified BSD License, compatible with Apache 2.0.