Starwrap

This is the python wrapper for Starspace. Which is not complete. Below are the APIs those are done and to be done.

done:

init
initFromTsv
initFromSavedModel
train
evaluate
getDocVector
nearestNeighbor
saveModel
saveModelTsv
loadBaseDocs
predictTags

to be done:

getNgramVector
printDoc
predictOne

How to build?

• make sure you have CMake installed. Otherwise install it from here
• install Conan c++ package manager from here.
• now, clone this repository, move to the directory StarSpace > python.
• run the build script.

chmod +x build.sh
./build.sh

• build script will download necessory packages, build wrapper and run test code.
• when it is done, you will find starwrap.so inside newly created build directory.
• you can either copy starwrap.so into your python project or set LD_LIBRARY_PATH (GNU/Linux) DYLD_LIBRARY_PATH (Mac OSX) and import is as a python module import starwrap.

How to use?

API is very easy and straightforward. Please refer test.py in test directory.

How to predict tags?

If you train a model in trainMode=0, you can get tag predictions. Here is the Python code to get prediction for barack obama in the political_social_media dataset.

sp.initFromSavedModel('tagged_model')
sp.initFromTsv('tagged_model.tsv')

dict_obj = sp.predictTags('barack obama', 10)
dict_obj = sorted( dict_obj.items(), key = itemgetter(1), reverse = True )

for tag, prob in dict_obj:
    print( tag, prob )

And you get this:

__label__obama 0.5291043519973755
__label__stopobamasamnesty 0.5073596239089966
__label__florida 0.5003609657287598
__label__entitlement 0.47724902629852295
__label__savesarah 0.4583539664745331
__label__keystone 0.4561977982521057
__label__1866 0.43746984004974365
__label__waroncoal 0.4283019006252289
__label__notleading 0.4117577075958252
__label__cuba 0.3887191414833069

For the full example, please refer to test_predictTags.py in test directory.