lightweight python wrapper for vowpal wabbit

View the Project on GitHub josephreisinger/vowpal_porpoise


Lightweight python wrapper for vowpal_wabbit.

Why: Scalable, blazingly fast machine learning.


  1. Install vowpal_wabbit. Clone and run make
  2. Install cython. pip install cython
  3. Clone vowpal_porpoise
  4. Run: python setup.py install to install.

Now can you do: import vowpal_porpoise from python.


Standard Interface

Linear regression with l1 penalty:

from vowpal_porpoise import VW

# Initialize the model
vw = VW(moniker='test',    # a name for the model
        passes=10,         # vw arg: passes
        loss='quadratic',  # vw arg: loss
        learning_rate=10,  # vw arg: learning_rate
        l1=0.01)           # vw arg: l1

# Inside the with training() block a vw process will be 
# open to communication
with vw.training():
    for instance in ['1 |big red square',\
                      '0 |small blue circle']:

    # here stdin will close
# here the vw process will have finished

# Inside the with predicting() block we can stream instances and 
# acquire their labels
with vw.predicting():
    for instance in ['1 |large burnt sienna rhombus',\
                      '0 |little teal oval']:

# Read the predictions like this:
predictions = list(vw.read_predictions_())

L-BFGS with a rank-5 approximation:

from vowpal_porpoise import VW

# Initialize the model
vw = VW(moniker='test_lda',  # a name for the model
        passes=10,           # vw arg: passes
        lbfgs=True,          # turn on lbfgs
        mem=5)               # lbfgs rank

Latent Dirichlet Allocation with 100 topics:

from vowpal_porpoise import VW

# Initialize the model
vw = VW(moniker='test_lda',  # a name for the model
        passes=10,           # vw arg: passes
        lda=100,             # turn on lda
        minibatch=100)       # set the minibatch size

Library Interace (TESTING)

Via the VW interface:

with vw.predicting_library():
    for instance in ['1 |large burnt sienna rhombus', \
                      '1 |little teal oval']:
        prediction = vw.push_instance(instance)

Now the predictions are returned directly to the parent process, rather than having to read from disk. See examples/example1.py for more details.

Alternatively you can use the raw library interface:

import vw_c
vw = vw_c.VW("--loss=quadratic --l1=0.01 -f model")
vw.learn("1 |this is a positive example")
vw.learn("0 |this is a negative example")

Currently does not support passes due to some limitations in the underlying vw C code.

Need more examples?


vowpal_wabbit is insanely fast and scalable. vowpal_porpoise is slower, but only during the initial training pass. Once the data has been properly cached it will idle while vowpal_wabbit does all the heavy lifting. Furthermore, vowpal_porpoise was designed to be lightweight and not to get in the way of vowpal_wabbit's scalability, e.g. it allows distributed learning via --nodes and does not require data to be batched in memory. In our research work we use vowpal_porpoise on an 80-node cluster running over multiple terabytes of data.

The main benefit of vowpal_porpoise is allowing rapid prototyping of new models and feature extractors. We found that we had been doing this in an ad-hoc way using python scripts to shuffle around massive gzipped text files, so we just closed the loop and made vowpal_wabbit a python library.

How it works

Wraps the vw binary in a subprocess and uses stdin to push data, temporary files to pull predictions. Why not use the prediction labels vw provides on stdout? It turns out that the python GIL basically makes streamining in and out of a process (even asynchronously) painfully difficult. If you know of a clever way to get around this, please email me. In other languages (e.g. in a forthcoming scala wrapper) this is not an issue.

Alternatively, you can use a pure api call (vw_c, wrapping libvw) for prediction.


Joseph Reisinger @josephreisinger



Apache 2.0