Refactor input handling to log inputs as artifacts for the current run. #71
Conversation
|
MLCommons CLA bot All contributors have signed the MLCommons CLA ✍️ ✅ |
src/modelplane/utils/input.py
Outdated
| pass | ||
| def log_artifact(self, current_run_id: str): | ||
| """Log the dataset to MLflow as an artifact for the given `current_run_id`.""" | ||
| mlflow.log_artifact(str(self.local_path()), run_id=current_run_id) |
There was a problem hiding this comment.
Why only log the local path? Can we also log the URL if it's a DVC input? Or the input is an artifact of a previous run?
There was a problem hiding this comment.
@bkorycki str(self.local_path()) is to provide it with the location of the file. The whole file is "logged" to mlflow, so it's visible/browse-able in that interface.
One thing we're losing is the details of the origin of the file, I could keep that as tags on the run maybe? Like an input tag with a nice string that represents the input?
There was a problem hiding this comment.
Ah got it. Yes, I think it would be nice to keep track of the provenance of these data files. Maybe log_input makes most sense for this?
There was a problem hiding this comment.
mlflow.log_input is just kind of annoying to deal with. (We had to have all that weird metadata stuff.) I feel like we could be much more human-friendly with tags?
I think we could log all the input params that we use to load the file. That might be simplest. Let me mock it up in a commit!
There was a problem hiding this comment.
Updated; results in something like this:
Have a look @bkorycki
…force MLflow active run requirement
Implements first two bullets of #68