Custom Charts: Slow performance when using ROC curve preset

A common use case I’d like to use wandb for is to store accuracy metrics across a number of experimental conditions and training bootstraps, and then compare those metrics faceted by experimental conditions, aggregating bootstraps as variance in the model performance.

I implemented a small demo of that in this dashboard, however the dashboard is incredibly slow for me; it takes 30s-1m to load the dashboard in a new tab, and normal interface usage like scrolling, clicking on panel elements, hovering, etc. often have a latency of 5+ seconds.

This is using wandb’s wandb.plot.roc_curve on a relatively small number of models and a relatively small test set (the ROC tables have 178 values in them each, which is less fine-grained than I’d normally like to see ROC metrics), so I’m concerned about using this in a real use case. Is this slowness expected for plots like this? How could I implement this behavior while retaining a snappy interface, or are there improvements that can be made to the web interface to support these use cases?

[cvphelps]

Thanks @stef-stripe, could you please shoot me an email at [e-mail] so we can talk next week?

[shawnlewis]

Hi folks, we just pushed a fix that improves the page behavior. Custom charts are new, there was an over-zealous rendering bug. There’s plenty more to optimize beyond this.

Can you try refreshing the pages and let us know if this first improvement helps?

As Carey said, we’ve love to talk on Monday.