feat: build temporary validation lists

[lessej]

Context

Attempts to build a validation list for ml labels. Takes an ML label, finds objects with that label, and builds a set of other labels on the objects. Some things to think about:

• Is there a way to make it more efficient? It's O(n^2) at least right now. Can some of the in-memory operations be offloaded to the DB query?
• What to do with labels that a user has added but then removed -- it seems like these are still on the object

[nathanielrindlaub]

@lessej what's the thinking with this line ^?

[nathanielrindlaub]

Can you remind me what you're trying to do here with there $seting of the Image.filteredLabels field?

[nathanielrindlaub]

This map should match the shape of the existing TARGET_CLASS array in the analysisConfig.js`, right? See example entry copied into the comment above.

[nathanielrindlaub]

I'm having a little trouble understanding this. Seems a little odd to use a reducer here. You're both dropping in validatingLabels as an initial value (132) then referencing it and manipulating the out-of-scope variable from within the reducer loop? But then returning the imageAcc after each run of the reducer function (131)? Then I kind of assume is what the reduce() returns and assigns imageAcc to validationLists once it's iterated over all the images? Maybe I'm not tracking or something but I think this could be accomplished with a simple forEach loop:

const validatingLabels = projectLabelIds.reduce((acc, lbl) => {
      return { ...acc, [lbl]: [] };
    }, {});

images.forEach((img) => {
      img.objects.forEach((obj) => {
        ...
         validatingLabels[firstValidLabelId] = new Set([...curr, ...labelIds]);
      }
}

[nathanielrindlaub]

I think this might be backwards. The firstValidLabel should be added to the array of validating labels, not used as the predicted label.

[nathanielrindlaub]

Just one small update and I think this is good to go!

[nathanielrindlaub]

We store all predicted labels above whatever the confidence threshold was set for that particular category, so we need to account for situations in which the model has returned two labels above the confidence threshold (i.e., more than one predicted label), rather than just grabbing the first.