slp 1.2.1

Speech, Language and Multimodal Processing models and utilities in PyTorch

slp

• Repo: https://github.com/georgepar/slp
• Documentation: https://georgepar.github.io/slp/latest/

slp is a framework for fast and reproducible development of multimodal models, with emphasis on NLP models.

It started as a collection of scripts and code I wrote / collected during my PhD and it evolves accordingly.

As such, the framework is opinionated and it follows a convention over configuration approach.

A heavy emphasis is put on:

• Enforcing best practices and reproducibility of experiments
• Making common things fast at the top-level and not having to go through extensive configuration options
• Remaining extendable. Extensions and modules for more use cases should be easy to add
• Out of the box extensive logging and experiment management
• Separating dirty / scratch code (at the script level) for quick changes and clean / polished code at the library level

This is currently in alpha release under active development, so things may break and new features will be added.

Dependencies

We use Pytorch (1.7) and the following libraries

• Pytorch Lightning
• huggingface/transformers
• Wandb
• Python 3.8

Installation

You can use slp as an external library by installing from PyPI with

pip install slp

Or you can clone it from github

git clone [e-mail]:georgepar/slp

We use poetry for dependency management

When you clone the repo run:

pip install poetry
poetry install

and a clean environment with all the dependencies will be created. You can access it with poetry shell.

Note: Wandb logging is enabled by default. You can either

• Create an account and run wandb login when you clone the repo in a new machine to store the results in the online managed environment
• Run wandb offline when you clone the repo to disable remote sync or use the --offline command line argument in your scripts
• Use one of their self-hosted solutions

Create a new project based on slp

You can use the template at https://github.com/georgepar/cookiecutter-pytorch-slp to create a new project based on slp

pip install cookiecutter poetry
cookiecutter gh:georgepar/cookiecutter-pytorch-slp
# Follow the interactive configuration and a new folder with the project name you provided will appear
cd $PROJECT_NAME
poetry install  # Installs slp and all other dependencies

And you are good to go. Follow the instructions in the README of the new project you created. Happy coding

Contributing

You are welcome to open issues / PRs with improvements and bug fixes.

Since this is mostly a personal project based around workflows and practices that work for me, I don't guarantee I will accept every change, but I'm always open to discussion.

If you are going to contribute, please use the pre-commit hooks under hooks, otherwise the PR will not go through the CI. And never, ever touch requirements.txt by hand, it will automatically be exported from poetry

cat <<EOT >> .git/hooks/pre-commit
#!/usr/bin/env bash

bash hooks/export-requirements-txt
bash hooks/checks
EOT

chmod +x .git/hooks/pre-commit  # Keep an up-to-date requirements.txt and run Linting, typechecking and tests

ln -s $(pwd)/hooks/commit-msg .git/hooks/commit-msg  # Sign-off your commit

Cite

If you use this code for your research, please include the following citation

@ONLINE {,
    author = "Georgios Paraskevopoulos",
    title  = "slp",
    year   = "2020",
    url    = "https://github.com/georgepar/slp"
}

Roadmap

• Optuna integration for hyperparameter tuning
• Add dataloaders for popular multimodal datasets
• Add multimodal architectures
• Add RIM, DNC and Kanerva machine implementations
• Write unit tests