sinr 0.1.4

Python implementation of graph community based word embeddings (SINr)

SINr

Build word embeddings based on community detection in graphs.

Project Organization

├── LICENSE
├── Makefile           <- Makefile with commands like `make data` or `make train`
├── README.md          <- The top-level README for developers using this project.
├── data
│   ├── external       <- Data from third party sources.
│   ├── interim        <- Intermediate data that has been transformed.
│   ├── processed      <- The final, canonical data sets for modeling.
│   └── raw            <- The original, immutable data dump.
│
├── docs               <- A default Sphinx project; see sphinx-doc.org for details
│
├── models             <- Trained and serialized models, model predictions, or model summaries
│
├── notebooks          <- Jupyter notebooks. Naming convention is a number (for ordering),
│                         the creator's initials, and a short `-` delimited description, e.g.
│                         `1.0-jqp-initial-data-exploration`.
│
├── references         <- Data dictionaries, manuals, and all other explanatory materials.
│
├── reports            <- Generated analysis as HTML, PDF, LaTeX, etc.
│   └── figures        <- Generated graphics and figures to be used in reporting
│
├── requirements.txt   <- The requirements file for reproducing the analysis environment, e.g.
│                         generated with `pip freeze > requirements.txt`
│
├── setup.py           <- makes project pip installable (pip install -e .) so src can be imported
├── src                <- Source code for use in this project.
│   ├── __init__.py    <- Makes src a Python module
│   │
│   ├── data           <- Scripts to download or generate data
│   │   └── make_dataset.py
│   │
│   ├── features       <- Scripts to turn raw data into features for modeling
│   │   └── build_features.py
│   │
│   ├── models         <- Scripts to train models and then use trained models to make
│   │   │                 predictions
│   │   ├── predict_model.py
│   │   └── train_model.py
│   │
│   └── visualization  <- Scripts to create exploratory and results oriented visualizations
│       └── visualize.py
│
└── tox.ini            <- tox file with settings for running tox; see tox.readthedocs.io

---

Background

Features

SINr is composed of two main modules :

• Cooccurrence : a cython based module to efficiently compute a cooccurrence matrix from a given corpus
• SINr : a module to compute cooccurence network based, sparse word embeddings

Installation

1. Launch a job on a slurm node -> srun -p gpu --gres "gpu:1" --time 1-0 --mem 5G --pty bash
2. Install conda
3. Clone repository -> git clone --branch nfm_sparse https://git-lium.univ-lemans.fr/tprouteau/sinr.git && cd sinr
4. Build conda environment -> conda env create -f environment.yml
5. Activate environment -> conda activate sinr_release
6. InstallSINr in development mode and SpaCy Transformer model for english -> cd src && python setup.py cythonize && pip install -e . && python -m spacy download en_core_web_trf
7. Use SINr!

Launch a Jupyter Notebook in jupyterlab

1. Activate your conda environment -> conda activate sinr_release
2. (upon first launch) install environment kernel in IPython -> ipython kernel install --name sinr_release --user
3. Launch a notebook on a node -> srun -p gpu --gres "gpu:1" --mem 80G -c15 -w "gpu15" jlaunch jupyter-lab #Use the -w option to choose the node one should not use a K20/K40 GPU as is it not supported by cupy anymore.
4. ctrl+click on the link displayed on the terminal and select the adequate kernel (sinr_release)

Usage

For additional examples see notebooks

Cooccurence

from sinr.cooccurrence import Cooccurrence
from sinr.pmi import pmi_filter

# Load your corpus as list of lists of tokens
sentences = [["sinr", "is", "fun"], ["sinr", "is", "a", "python", "package"]]
# Build cooccurrence matrix
c = Cooccurrence()
c.fit(sentences, window=2)

#Normalise cooccurrence matrix using PPMI
c.matrix = pmi_filter(c.matrix)
c.save("/path_to_output/matrix.pk")

SINr

The extraction of the embedding is currently greedy in terms of memory. When working with large corpora, do not hesitate to ask for rather large amounts of RAM (>100G)... This is currently being fixed.

from sinr.graph_embeddings import SINr

model = SINr.sinr("/path_to_output/matrix.pickle", output_path="path_to_output", n_jobs=4)  
#If an output_path is supplied, the model will be saved -- Embeddings are returned
#as a Model object comprised of a dictionnary for the vocabulary and a scipy.sparce.csr_matrix for the vectors

Contributing

Pull requests are welcome. For major changes, please open an issue first to disccuss the changes to be made.

Compile/Install from source

In order to compile and install SINr from source follow the procedure described below

git clone --branch nfm_sparse https://git-lium.univ-lemans.fr/tprouteau/sinr.git
cd sinr
conda env create -f environment.yml
conda activate sinr_release
python setup.py cythonize
pip install -e .

Evaluate Word Embeddings

In order to evaluate the word embeddings on the similarity task you may use the library Word Embedding Benchmarks developped by Stanislaw Jastrzebski : https://github.com/kudkudak/word-embeddings-benchmarks

⚠️ The embeddings returned by the model are of type Scipy.sparse.csr_matrix you will need to pass them as a dense matrix using the function

matrix = my_sparse_csr_matrix.todense()

Refer to the documentation and examples to know which format to use in input of the benchmarking library.

License

Project based on the