nav[aria-label="Primary Navigation"] { padding: 0; & ul { list-style: none; width: 100%; display: flex; flex-direction: row; justify-content: start; align-items: start; gap: 30px; padding: 0; & li { margin: 0; } & ul li { list-style: none; } } }

Examples of model deployment via Formula Depot to show customers?

Ben_BarrIngh

Hi All,

I have a number of customers interested in JMP Pro for developing models for deployment into their manufacturing environments - I've had a few ask for an example of how the models from the Formula Depot can be applied externally. Do we have any 'easy' demos that can show the models being used in a different environment (like Python?).

Cheers,

Ben

Accepted answers

Laura_Higgins

Hi Ben - Frequently just showing people the code that JMP produces out of the Formula Depot is sufficient. In particular, showing the C and Python code usually does it for people. Don't get pulled into supporting things outside of JMP like code deployment into their systems. That's for them to figure out.

A nice demo of Formula Depot includes model comparison and the Profilers (although Model Screening does this now as well, and you can put the formulas into the stand alone Profiler under the Graph menu). I like to make 3 models of different types, such as a Response Surface model, a Partition model, and either a Bootstrap or Neural model. When they are stacked up in the Profiler, you can show people that they can understand how each kind of model is related to the data, and each model is different ("Look at how this one is stepped, this one is a curved line, etc."). Also show the Actual by Predicted, and how some areas of the data fit differently than others in different models. Yield from White Polymer works for this.

All comments

russ_wolfinger

Hi @Ben_BarrIngh , This topic is being actively discussed by Product Management in conjunction with ways to improve and move beyond Formula Depot.

Wanted to offer a positive direction we have been using with the XGBoost and Torch add-ins, which is to train models with no coding in JMP Pro and then deploy them in Python using custom wrapper Python classes to read the models saved by JMP Pro back into Python.

More generally, I believe practically every model JMP produces could be readily implemented in PyTorch, building on Python code already available from XGBoost and Torch. This would provide feature-rich deployable Python model objects runnable on GPUs and convertable to other popular formats like ONNX.

If you have one or more specific examples in mind--or maybe even a customer willing to recommend or provide one from a manufacturing environment--I would be happy to work with you to get formulas currently produced by JMP converted into full-fledged PyTorch model objects.

Simple regression models (linear or logistic) would likely be the easiest. One key part is ensuring new X variables get transformed in exactly the same way the JMP model does them.

Laura_Higgins

Ben_BarrIngh

Hi @Laura_Higgins,

Thanks for the advise and the guidance on how to demo the Formula Depot - will give it a try!

Thanks,

Ben

Ben_BarrIngh

Hi @russ_wolfinger ,

Thanks for the information - I'll have a look into some datasets that may be suitable - generally looking for something larger with a manufacturing angle that would cover most customer requirements and uses a simple model - will let you know!

Thanks,

Ben