Recently we had some breaking changes to eval_single_step.py merged to main due to insufficient test coverage. This PR extends the coverage by adding tests for compute_metrics (using both a forward and a backward reaction model). It also improves error handling in the evaluation script (errors otherwise produced by OmegaConf tend to be a bit cryptic) and allows get_model to make use of default model weights (otherwise this feature was not accessible when running through one of the the CLI entrypoints).
Recently we had some breaking changes to
eval_single_step.py
merged tomain
due to insufficient test coverage. This PR extends the coverage by adding tests forcompute_metrics
(using both a forward and a backward reaction model). It also improves error handling in the evaluation script (errors otherwise produced byOmegaConf
tend to be a bit cryptic) and allowsget_model
to make use of default model weights (otherwise this feature was not accessible when running through one of the the CLI entrypoints).