Feature request: allow trained model to be used within PL

Goal: extend the data wizardry to allow PL to accept a second datafile (necessarily identical in structure etc. to the initial, e.g. CSV) data file, run the model against the data and

a) create outputs (images, other) as appropriate by running the trained model against this file
b) duplicate the second data file and save into it:
b.1) model outputs - paths to image outputs, values for e.g. logistic predictions etc., labels…
b.2) additional info (e.g. object detection rectangles, probabilities…)

Rationale: currently a trained model has to be saved and served by a TF server, this creates an extra barrier to deriving value from PL’s visual modelling capabilities. The technical skills/resource access required to deploy a model are not at the same level as those for using PL to build a model - but what is the value of building a model - other than learning how - if it cannot easily be used?

NB “models” are idealisations for particular purposes… it’s only when they are used in real life that their strengths and weaknesses (e.g. hidden assumptions) become readily apparent. Smoothing the path to using PL models would also enable iterative development by users… and their shareability :wink:


Hi @JulianSMoore,
Thanks for the request, and great point with the inference! :slight_smile:

I think we might do it in two steps, where the first step is to provide a script together with the model, where running the script would be to run the model (or serve the model).
The second step would then be what you suggested, where the inference can happen straight in PL. Although we’ll see if we can’t jump straight to step 2 as it’s not a huge amount of extra effort.

1 Like