PyTorch implementation of StarSpace

Based on the approach described in StarSpace: Embed All The Things! by Ledell Wu, Adam Fisch, Sumit Chopra, Keith Adams, Antoine Bordes and Jason Weston

NOTE: The current version is work in progress and doesn't yet match the functionality of the original implementation.

The C++ version maintained and developed by the authors can be found here.

Installation

Tested on Pyton 3.6.1 and PyTorch 0.3.1

• Install PyTorch via conda or pip

• Install dependencies
  $ pip install -r requirements.txt

Usage

The commandline interface resembles the one used by the C++ version.

To train and validate a StarSpace model on the AG news corpus, run the following commands:

Train a model

$ ./starspace train 
    --train-file "<DATSET_PATH>/train.csv" 
    --model-path "<MODELDIR>/" 
    --file-format ag_news 
    --d-embed=10 
    --lr=0.01 
    --epochs=5

Validate the trained model

$ ./starspace test 
    --test-file "<DATSET_PATH>/test.csv" 
    --model-path "<MODELDIR>/" 
    --file-format ag_news