Environment Setup

It is recommended to create a new Python virtual environment for running anemoi-inference to isolate tasks within a ML workflow. When creating this environment it is also recommended to ensure that the versions of key packages are compatible / identical to those used during training.

This is of particular importance for the following packages:

• anemoi-models

• anemoi-graphs

• torch

• torch_geometric

Note

If using the dataset input source, you will need anemoi-datasets installed.

Important

As anemoi is still in active development, it is recommended to use the same version of the above anemoi packages as those used during training.

Tip

You can check the versions of the packages used during training by inspecting the checkpoint metadata with the inspect command and getting a list of requirements:

anemoi-inference inspect --requirements /path/to/inference-last.ckpt

See also

• Installing - Installation instructions

• Inspect Command - Inspect checkpoint metadata and requirements

• Quickstart - Quickstart guide

• Optimisation and Performance - Performance optimisation tips