Acumen ✨ Template ✨

Before starting any deep learning research project, use this template to make your life as a researcher easier.

Keep your sanity! Work with wandb (included by default)!

Coded with love and coffee ☕ by Adrian Cosma. But I need more coffee!

Open Source projects that used ✨ The Template ✨

• The Strawberry Problem 🍓: Emergence of Character-level Understanding in Tokenized Language Models
• How Hard is this Test Set? NLI Characterization by Exploiting Training Dynamics
• Reading Between the 🎞️ Frames: Multi-Modal Depression Detection in Videos from Non-Verbal Cues
• It’s Just a Matter of Time: Detecting Depression with Time-Enriched Multimodal Transformers
• Learning Gait Representations with Noisy Multi-Task Learning
• Exploring Self-Supervised Vision Transformers for Gait Recognition in the Wild

Usage:

Install cookiecutter with: conda install cookiecutter

Create directory for project structure with:

cookiecutter gh:cosmaadrian/acumen-template

Main Philosophy

All deep learning projects have three main components: a dataset, a model, a training procedure and an evaluation procedure.

This project's goal is to provide a ready-made deep learning research project structure. This template / framework is opinionated, assumes the usage of pytorch and wandb.

It is heavily based on .yaml configuration files and command line arguments to run experiments in a declarative way. The only things that the researcher needs to focus their time and energy is "bussiness logic", which means the data processing / model training / model evaluation. The main training loop and experiment tracking is automatically done by the framework. The usual callbacks for model checkpointing, early stopping and logging are provided by default.

Command line management: 🔧Forge

You can use the lib/forge.py command line tool to better manage your project. Get started:

python lib/forge.py help

Adding classes for Datasets, Models, Trainers and Evaluators using 🔧Forge

Datasets, Models and Trainers must be placed in the appropriate directory, added to their respective __init__.py file, and added to the nomenclature.py file.

Structure for the configuration file.

You'll get the hang of it!

Extending configs.

Use $extends$: <path> in a config file if you want to extend it.

configs/a.yaml foo: bar

configs/b.yaml $extends$: configs/a.yaml baz: zab Running with this config file will result in args == {'foo': bar, 'baz': zab}

Including other configs

Use $includes$: [<path1>, <path2>, ...] to include other configuration files directly. configs/a.yaml foo: bar

configs/b.yaml $includes$: - configs/a.yaml Running with this config file will result in args == {'foo': bar}

Referencing other values

You can reference another value from the current configuration using the value = ${other_value} syntax. a: b c: ${a} Running with this config file will result in args == {'a': b, 'c': b}

Update to the latest version

To keep your project updated with the latest lib/ folder, just run the command:

python lib/forge.py update

This command will override the lib folder with the latest changes. WARNING. Any changes you made to lib/ will be lost.

🎓 Citation

If you used this template in your projects, please cite this repository:

@software{cosma23acumen, author = {Cosma, Adrian}, doi = {10.5281/zenodo.8356189}, month = {9}, title = {{AcumenTemplate}}, url = {https://github.com/cosmaadrian/acumen-template}, version = {0.1.1}, year = {2023} }