Superfast CUDA implementation of Word2Vec and Latent Dirichlet Allocation (LDA)

This project is to speed up various ML models (e.g. topic modeling, word embedding, etc) by CUDA. It would be nice to think of it as gensim's GPU version project. As a starting step, I implemented the most widely used word embedding model, the word2vec model, and the most representative topic model, the LDA (Latent Dirichlet Allocation) model.

• Python3.6+
• gcc / g++ (>= 5.1 for c++14)
• cuda >= 7.0
• Tested on Ubuntu 18.04 / GCC 7.5 / CUDA 11.1 / Python 3.6

• install from pypi

pip install cusim

• install from source

# clone repo and submodules
git clone git@github.com:js1010/cusim.git && cd cusim && git submodule update --init
# install requirements
pip install -r requirements.txt
# generate proto
python -m grpc_tools.protoc --python_out cusim/ --proto_path cusim/proto/ config.proto
# install
python setup.py install

• examples/example_w2v.py, examples/example_lda.py and examples/README.md will be very helpful to understand the usage.
• paremeter description can be seen in cusim/proto/config.proto

• AWS g4dn 2xlarge instance is used to the experiment. (One NVIDIA T4 GPU with 8 vcpus, Intel(R) Xeon(R) Platinum 8259CL CPU @ 2.50GHz)
• results can be reproduced by simply running examples/example_w2v.py and examples/example_lda.py
• To evaluate w2v model, I used evaluate_word_pairs function (ref link) in gensim, note that better performance on WS-353 test set does not necessarily mean that the model will workbetter in application as desribed on the link. However, it is good to be measured quantitively and fast training time will be at least very objective measure of the performaance.
  • I trained W2V model on quora-duplicate-questions dataset from gensim downloader api on GPU with cusim and compare the performance (both speed and model quality) with gensim.
• To evaluate LDA model, I found there is no good way to measure the quality of traing results quantitatively. But we can check the model by looking at the top words of each topic. Also, we can compare the training time quantitatively.
• W2V (skip gram, hierarchical softmax)

• W2V (skip gram, negative sampling)

• W2V (CBOW, hierarchical softmax)

• W2V (CBOW, negative sampling)

• LDA (nytimes dataset from https://archive.ics.uci.edu/ml/datasets/bag+of+words)
  • I found that setting workers variable in gensim LdaMulticore does not work properly (it uses all cores in instance anyway), so I just compared the speed between cusim with single GPU and gensim with 8 vcpus.
  • One can compare the quality of modeling by looking at examples/cusim.topics.txt and examples/gensim.topics.txt.

• support half precision
• support multi device (multi device implementation on LDA model will not be that hard, while multi device training on w2v may require some considerations)
• implement other models such as FastText, BERT, etc
• contribution is always welcome