I am using DVC to store some models but the way I am training my models does not allow me to easily use the pipelines. And it does not necessarily make sense in my case. Nevertheless, I have some metadata I would like to keep as well as some metrics.
So my question is how should I do this? Because reading the doc and looking at some posts form here, I should use the dvc.yaml. But cmd is not optional and does not really make sense in my case. So right now I could just have a fake cmd but this does not sounds right.
I am using the management features already. What I would like is to be able to do is to define somewhere a set of metadata to track my progress. So exactly metrics and metadata from the dvc.yaml would fit the bill. But I would like to be able to do this without the need to run a package.
Let’s say that when I get a new model, I don’t have control over the training or testing process. I just integrate the models and want to track them. So when I get a new model, I get it with a set of metadata and metrics. And I want to track these.
But metrics and plots subcommands can be used with any file in the repo. So you can version the metadata with Git directly (along with your .dvc files) and use dvc metrics diff, dvc plots show, etc. on any supported files (YAML, TOML, CSV, TSV).
See the optional [targets] argument or sometimes --targets flag in each command.
Oh nice indeed! IMO this could be highlighted in the doc that we can use these features without relying on dvc.yaml. Like in the use case versioning or in the user guide. (Open to make a PR or at least an issue if you want)
So now I guess I could put my metadata either in a separate file or in params.
!