Supportr

Intro

supportr is a package used to predict the value of support of texts.

It is based on a fine tuned BERT model.

Install

Use pip

If pip is installed, supportr could be installed directly from it:

pip install supportr

Dependencies

python>=3.6.0
torch>=0.4.1
numpy
pandas
unidecode
pytorch-pretrained-bert
pytorch-transformers

Usage and Example

Notes: During your first usage, the package will download a model file automatically, which is about 400MB.

predict

predict is the core method of this package, which takes a single text of a list of texts, and returns a list of raw values in [1,5] (higher means more support, while lower means less).

Simplest usage

You may directly import supportr and use the default predict method, e.g.:

>>> import supportr
>>> supportr.predict(["I am totally agree with you"])
[3.8364935]

Construct from class

Alternatively, you may also construct the object from class, where you could customize the model path and device:

>>> from supportr import Supportr
>>> sr = Supportr()

# Predict a single text
>>> sr.predict(["I am totally agree with you"])
[3.8364935]

# Predict a list of texts
>>> preds = sr.predict(['I am totally agree with you','I hate you'])
>>> f"Raw values are {preds}"
[3.836493  1.7458204]

More detail on how to construct the object is available in docstrings.

Model using multiprocessing when preprocessing a large dataset into BERT input features

If you want to use several cpu cores via multiprocessing while preprocessing a large dataset, you may construct the object via

>>> pr = Supportr(CPU_COUNT=cpu_cpunt, CHUNKSIZE=chunksize)

If you want to faster the code through multi gpus, you may construct the object via

>>> pr = Supportr(is_paralleled=True, BATCH_SIZE = batch_size)

Contact

Junjie Wu (wujj38@mail2.sysu.edu.cn)