[AiLab] correctly access data in testValues #39097
Conversation
apps/src/MLTrainers.js
Outdated
| @@ -48,7 +48,7 @@ export function predict(modelData) { | |||
| convertTestValue( | |||
| modelData.featureNumberKey, | |||
| feature, | |||
| modelData.testData[feature] | |||
| modelData.testData[feature.replace(/\W/g, '')] | |||
There was a problem hiding this comment.
Since we have this snippet a few times now, can we hoist it to a helper function?
There was a problem hiding this comment.
Where should it live? It's used in both ai.js and here in MLTrainers.js. Just pick one and export to the other?
There was a problem hiding this comment.
I'm not too sure about the /apps heirarchy, but I think it makes sense for applab/ai.js to reach into its parent directory for such a utility function (since it seems we have a variety of "utils" and "common" files there). It might be cleanest to add this to apps/src/utils.js so that we don't create a dependency between the two existing source files, or to create a new apps/src/aiUtils.js file?
There was a problem hiding this comment.
Sounds good. I went with the aiUtils option.
tldr: "Whipped Cream" != "WhippedCream"
We got reports from the curriculum team of a bug where the prediction in AI Lab didn't match the prediction in App Lab for the same trained model with the same test values. This initially seemed very worrisome, but was actually quite a straightforward fix.
In #39067 we striped column names from AI Lab trained models of whitespace and special characters so they could be used as design element ids and keys in the
testValuesobject.However, in the
predictfunction we were not correctly accessing values intestValuesto pass to theconvertTestValuefunction from #39044. Therefore, certain features (particularly those with spaces, e.g. "whipped cream" and "fried chicken") were being converted to NaN, which was causing the trained model to return inaccurate predictions.BEFORE:
AFTER:
Now, predictions are consistent across AI Lab and App Lab.