Open mnrozhkov opened 9 months ago
Another thought on a lightweight way to help here: better docs in https://dvc.org/doc/dvclive/live/log_plot. Having an example of the input format for each template could go a long way. There are already examples of different templates in https://dvc.org/doc/command-reference/plots/show that we could use as a starting point.
Background on the current implementation: https://github.com/iterative/dvclive/pull/543#pullrequestreview-1402602708
Marking as p2 since I don't think log_plot()
is frequently used, but still would be really nice to have these improvements
When people start using
Live.log_plot()
, they could struggle with getting an expected visualization because of 2 reasonslog_plot()
is very opinionated about the data format required for every templateHere are some ideas to help with DVCLive onboarding:
1. "Relax" requirements for data formats supported
For example, the
bar_horizontal
template expects smth like this:It would be cool to support other formats like: 1) Pandas DataFrame
2) Dict with automatically extracts keys as
y' and values as
x.`2. Provide minimal sanity checks for data/configs provides For example, if I run this code snippet:
I'll not get any error, but there is nothing showing in VSCode after that:
Reason? There is a mistake in
x
andy
arguments assignment, the correct isy="name", x="importance"
. But, it's very easy to oversee this typo and spend a lot of time trying to figure it out.How can we help?
check that the
bar_horizontal
template expects numerical data forx
3. Provide good warning messages and hints if formats incompatible If we have data/args checks, we may tell about this in warning messages and this will help a lot to see smth like: