Environment Management

We try to maintain a central code repository for $\Psi_0$ as well as all the baselines.

Key considerations are:

• We want to train each baseline with minimum code change, so we create separate isolated environment for each baseline.
• At the same time, we want all the baseline models, dataloaders, etc can be mixed in the same environment.

So we manage the environments as illuated below:

To conclude:

1. $\Psi_0$ is managed using the default pyproject.toml, and depdendency groups include base, psi, viz and serve.
2. Each baseline install src/ package and base dependencies, so all the baseline models, dataloaders etc., can be used in the envrionment.
3. Each baseline maintain a seperate requirements.txt which include the baseline's torch and transformers versions.