Compare different architectures for a same task

Hey there.

I discovered W&B recently and decided to use it for my current research project. The thing is that I have a specific task to solve and would like to evaluate a bunch of completely different model architectures, having different sets of hyper-parameters.

Most of the online resources and tutorials I’ve found only shows examples of W&B usage to evaluate different experiments with different params selections (e.g. optimized using sweeps). However none of the examples I found explained how to optimally organize a W&B project including different architectures to solve the same task, and thus being able to compare in a glimpse the different performances in a single view / report.

My idea was to make use of the job_type flag and group every architecture instances together under a same job_type flag. But still seems like not the best solution and was wondering if there is some special feature or built-in tool that I’ve not noticed yet (or even good practices?).

(Other than that, W&B looks really insane).

Many thanks :smile:

Hey @andreakiro, apologies abou the late response on this. Grouping is the solution I’d suggest here. You can pass a group argument to wandb.init. After that you’ll be able to filter the dashboard based on the architecture you are interested. You can also use tags for this purpose.

This topic was automatically closed 60 days after the last reply. New replies are no longer allowed.