A python based RPC server that encapsulates the functionality of Vowpal Wabbit making it easier to integrate it into your data workflow. The RPC protocol is language agnostic so you can access from any kind of platform.
sudo pip install git+git://github.com/prashanthellina/vwserver.gitStart a VWServer by doing. Note that VWServer will use the specified directory "data" for storing all of its data. Make sure you do not have any of your own data there. You can change the port number to anything else you desire.
vwserver run --port 8889 data `which vw`Now the server is running and is web accessible. Open http://localhost:8889 in the browser. You will see a web-based python console to interact with this VWServer.
Try running the following commands and observe the output
>>> dir()
['__builtins__', 'api', 'call', 'logging', 'server']
>>> help(api)
Help on VWAPI in module vwserver.vwserver object:
class VWAPI(__builtin__.object)
| Methods defined here:
|
| __init__(self, data_dir)
|
| create(self, name, options=None)
| Creates a new VW model with @name and using @options
|
| destroy(self, vw)
| Destroy the specified VW model from both memory
| and disk permanently.
|
| predict(self, vw, items)
| Perform prediction using @vw model on the provided @items.
| @items - a list of strings representing the input lines
| in the VW format
|
| returns: a list of response lines as returned by VW
|
| save(self, vw)
| Saves the model learnt so far
|
| show_options(self)
| Shows the allowed options and their default values
|
| shutdown(self)
| Stop the server
|
| train(self, vw, examples)
| Train the @vw model using @examples
| @examples - a list of strings representing example lines
| in the VW format
|
| returns: a list of response lines as returned by VW
|
| unload(self, vw)
| Unloads a VW model from memory. This does not
| destroy the model from disk and so it can be
| loaded again later for usage.
|
| ----------------------------------------------------------------------
| Data descriptors defined here:
|
| __dict__
| dictionary for instance variables (if defined)
|
| __weakref__
| list of weak references to the object (if defined)Now let us check if there are any models currently loaded (We expect to see none as we are just starting off)
>>> api.vws()
{}Let us create a new Vowpal Wabbit model.
>>> api.create('foobar')And that's it! We are ready to use this for training and prediction. Try it out - You can refer to the image at the top of this document for a reference on how to perform training and prediction.
Note that we can pass options while creating a new VW model using the create method. Here is an example.
>>> api.create('foobar1', {'bit_precision': 20})If you want to see what options are available and their default values, do
>>> api.show_options()These options are directly converted into command line parameters for the vw command. Let us now clean-up by destroying both the models we created.
>>> api.destroy('foobar')
>>> api.destroy('foobar1')In the document so far we have seen how to interact with the VWServer using it's web based python console. However for using it in a practical setting we will need a mechanism to access it remotely from another application. VWServer is built on top of FuncServer and hence supports having an API accessible over the network. Funcserver is a lightweight tool - you should check out it's documentation and code to get a better idea on how to interact with VWServer.
Here is an example python script: example_client.py
from vwserver import VWClient
vw = VWClient("http://localhost:8889")
vw.create('foobar')
# do what every training and prediction just like in the web based python console
vw.destroy('foobar')
# we are destroying only to cleanup as a part of this example. In a real scenario, you would want to keep the model!