Dewey
Dewey is a fast reproducible training automation tool for MLOps pipelines.
Dewey is a machine learning automation tool written to create consistent reproducible ways to train models in a framework agnostic way. It allows providing a training specification, and the Dewey training framework takes care of all of the standard boilerplate code involving writing training loops, monitoring & metrics, managing model checkpoints, and more.
Please note: Don't expect much... YET! This repository is currently in alpha development. It is subject to rapid updates and breaking API changes. But please stay posted. Good stuff is on the way.
Usage
Install using pip:
pip install pydewey
Create a file called train.py
. Define your hyperparameters, model, loss, and optimizer, and Dew(ey) it up real good!
For an example, put the train.py
file in this repo's examples
folder in your current directory, and run the command dwy
.
Todos:
- RNG seed handling for train.py and for actual training/ reproducibility
- plugin deps on other plugins (install once)
- build on above for full plugins for each framework
- plugin priorities?
- model naming conventions for checkpoints
- distributed training
- better support multi-model workflows
- test
- optimize
- doc