Hello, I am trying out

kedro-mlflow

and I have problems tracking a simple dict using

MlflowMetricsDataSet

in my catalog. It seems to me that

mlflow.log_metrics()

accepts a normal dict, but the data expected by

MlflowMetricsDataSet

needs a different format (I get an exception that a key

value

doesn't exist). Does anyone have experience using it?

Hi @Vassilis Kalofolias, actually this was originally by design to ensure

save

and

load

method will handle the same object. TBH I think nobody loads metrics from mlflow with this dataset, and it will make the API simpler and more consistent with mlflow to be able to log simply a

dict

. You can open an issue so I'll keep track of it and fix it one day, but it is quite low priority so if you want to change thing soon, feel free to open a PR! The good news is that

MlflowMetricDataSet

can handle this, so you can return a tuple at the end of a node instead of a dict and it will work except if you have dynamic names in your metric dict, since you have to write a catalog entry for each of them

Hi, thanks a lot! I agree about the consistency with MLflow part, also that saving is probably enough. I will try

MlflowMetricDataSet

and eventually open a ticket.