This repository contains a fast, scalable, highly-parallel Python implementation of the GloVe [1] algorithm for word embeddings (found in solvers.py) as well as code and scripts to recreate downstream-task results for unsupervised DisC embedding paper. An overview of the latter is provided in this blog post at OffConvex.
If you find this code useful please cite the following:
@inproceedings{arora2018sensing,
title={A Compressed Sensing View of Unsupervised Text Embeddings, Bag-of-n-Grams, and LSTMs},
author={Arora, Sanjeev and Khodak, Mikhail and Saunshi, Nikunj and Vodrahalli, Kiran},
booktitle={Proceedings of the 6th International Conference on Learning Representations (ICLR)},
year={2018}
}An implementation of the GloVe optimization algorithm (as well as code to build the vocab and cooccurrence files, optimize the related SN objective [2], and optimize a source-regularized objective for domain adaptation) can be found in solvers.py. The code scales to an arbitrary number of processors with virtually no memory/communication overhead. In terms of problem size the code scales linearly in time and memory complexity with the number of nonzero entries in the (sparse) cooccurrence matrix.
On a 32-core computer, 25 epochs of AdaGrad run in 3.8 hours on Wikipedia cooccurrences with vocab size ~80K. The original C implementation runs in 2.8 hours on 32 cores. We also implement the option to use regular SGD, which requires about twice as many iterations to reach the same loss; however, the per-iteration complexity is much lower, and on the same 32-core computer 50 epochs finish in 2.0 hours.
Note that our code takes as input an upper-triangular, zero-indexed cooccurrence matrix rather than the full, one-indexed cooccurrence matrix used by the original GloVe code. To convert to our (more disk-memory efficient) version you can use the method reformat_coocfile in solvers.py. We also allow direct, parallel computation of the vocab and cooccurrence files.
Dependencies: numpy, numba, SharedArray
Optional: h5py, mpi4py*, scipy, scikit-learn
* required for parallelism; MPI can be easily installed on Linux, Mac, and Windows Subsystem for Linux
Scripts to recreate the results in the paper are provided in the directory scripts-AKSV2018. 1600-dimensional GloVe embeddings trained on the Amazon Product Corpus [3] are provided here.
Dependencies: nltk, numpy, scipy, scikit-learn
Optional: tensorflow
[1] Pennington et al., "GloVe: Global Vectors for Word Representation," EMNLP, 2014.
[2] Arora et al., "A Latent Variable Model Approach to PMI-based Word Embeddings," TACL, 2016.
[3] McAuley et al., "Inferring networks of substitutable and complementary products," KDD, 2015.