Combination explorer

Pick a method, swap a stage, and see whether it still trains. Every answer is a recorded probe, a stored answer of grasppanda compose, or says that nothing was measured. On your machine, grasppanda explore --weights-dir <dir> also runs a ported method on a scene, in this page.

Questions

Where do the answers come from?

From records, never from a guess. A verdict is what grasppanda compose and grasppanda probe gave for exactly the combination shown. Without a record, the page may hold the answer of compose alone, drawn with a dashed outline: it says the stages fit, not that the combination trains. It is stored for one change and for two changes of two slots; a stage together with a recipe option has no stored compose answer. Anything else is not measured, and the page prints the two commands that find out. How each level is judged: Evidence.

Does a green tile mean a good combination?

A probe shows that a combination trains on real frames of GraspNet-1Billion, not how good it is. Accuracy is claimed only for a method trained and evaluated here.

No accuracy appears on this page. Trains like shipped goes one step further: a swap that ran was trained longer, beside the component the method ships, and their losses were compared. How we measure.

Why does a method offer these options?

A slot lists what grasppanda compose <method config> --list <slot> lists for that method, and the method's page shows the same list. A method without a slot says so in Compare: HGGD reads an RGB-D image and has an image encoder instead of a scene encoder.

Can I run a method here?

On your machine. grasppanda explore serves this page from the installed package and checks a selection of any number of slots with your own compose, saves it as a config, and never trains: the probe stays a command you run. With weights, grasppanda explore --weights-dir <dir> also enables the Run tab: pick a method and a weights file, and run it on the example scene or on your depth image (Run in the browser).