Dear W&B Community,
I have system metrics logged like the “time per step” or “time per backward pass” for a model.
When doing this on different hardware, I would like to compare the effect this has on these metrics.
In the following examples, I profile the basic Torch CIFAR10 model on a 1,2,4,8,16 and 32 CPU VM.
When looking at a Linechart
, the full history of these metrics is visible, however, it is very hard to compare them due to the overlapping and oscillation:
When using a Barchart
, only the last value is visualized:
The functionality that would be nice is to group values based on their count or occurrence, as grouping by runs already works perfectly. Here’s the same data but run through seaborn.barplot
:
Would this be possible to implement? Or does anybody know a way to get that functionality?
My current workaround is to download the data manually and run it through seaborn. Unfortunately, I did not understand the errors I’ve gotten with the Custom Chart
functionality when trying to port Vega examples to use wandb as a data basis.
I’d be very glad if anybody can point me to a tutorial on how to migrate existing Vega examples to be used with wandb (and the common problems, like differences between v3/v4/v5, as these seemed to be an issue for me).