log_plot: allow for y to be a list

[dberenbaum]

Should we allow for y to be a list?

Edit: can be a follow up, so not blocking, but I think it makes it significantly more useful.

Originally posted by @dberenbaum in #543 (comment)

[jenfredwell]

+1 for this. I came here looking for a way to log a plot with a definition like

- data.json:
    template: linear
    x: c
    y:
    - a
    - b

Is this possible to write directly to data.json instead of live.log_plot(..) and have this definition in the dvc.yaml file?

[dberenbaum]

Is this possible to write directly to data.json instead of live.log_plot(..) and have this definition in the dvc.yaml file?

Yes. You can always manually write to json and configure dvc.yaml yourself. See https://dvc.org/doc/user-guide/experiment-management/visualizing-plots for a guide on how to do it.

[jenfredwell]

Thanks. I was able to add a plot like this to the dvc.yaml, but it doesn't seem to appear in DVC Studio until a git commit is performed. I was hoping it would be updated live. Is this expected behavior?

[dberenbaum]

Yes, live updates are limited to dvclive plots. You should be able to still see the manual plots once the experiment completes if you run dvc exp push. Would that be enough, or is it important to have the plots updating in realtime?

[jenfredwell]

I see. Thanks for the clarification. I think that should be fine, but would be nice to get live updates. It's just not what I expected.

[dberenbaum]

@jenfredwell Have you actually tried it? I think it will work even though it's not documented. See this example:

import pandas as pd
from dvclive import Live
from sklearn.datasets import load_iris

iris = load_iris()
datapoints = pd.DataFrame(data=iris.data, columns=iris.feature_names)

with Live() as live:
    live.log_plot(
        "sepal",
        datapoints,
        x="sepal length (cm)",
        y=["sepal width (cm)", "petal width (cm)"],
        template="scatter",
        title="Sepal length vs Sepal/petal width"
    )

This renders a plot like:

If you confirm it works for you, we can mention support in the docs as well.

[jenfredwell]

Yes, this actually does work but doesn't seem to update in real time. It looks like the plots/custom directory doesn't get created until the experiment is over.

[dberenbaum]

Sorry @jenfredwell, I had forgotten that Studio does not support live updates of log_plot() results at the moment.

[jenfredwell]

Ok good to know. Is this in the roadmap by chance? It would be helpful to us to consolidate some of what we're starting to track with dvclive.

[dberenbaum]

Sorry, it's not in the roadmap at the moment