angryrou / udao Goto Github PK

Petry is a modern way to manage dependencies and package code.
It will avoid having two sources for dependencies: removing the need for requirements.txt and only keeping pyproject.toml

https://python-poetry.org/docs/

Optimization algorithms require independent functions for each objective or constraint, impacting optimization speed, which may not match the speed achieved in our referenced papers.

E.g., if the cost function is CPU * hour, we need to compute latency twice to get a 2-d objective tuple (latency, cost). But before, we calculated cost based on the latency results and reduced one latency model call in each iteration.

Another example is that when a user trains a multi-task model and can provide multiple objectives in one path. The current optimizer requires to call the model twice for 2 objectives bc it always considers objective functions to be independent.

To improve this, we will consider

1. Construct a class, say MooModel to wrap functions from all optimization elements (objectives and constraints). So the user has the flexibility to reuse the result from latency to construct the cost to avoid duplicated computation.
2. Refactor SOProblem/MOProblem and SOO solvers to adapt the MooModel on how to derive objective values.

Our test workflow takes 1-2 min to install all pip dependencies.

1. can we add the pip cache as https://github.com/actions/setup-python#caching-packages-dependencies

2. is there any best practice to follow when updating the requirements.txt?