Handle large mutliclass better in WIT #2385
Conversation
|
@tolga-b please review |
tolga-b
left a comment
tolga-b
left a comment
There was a problem hiding this comment.
LGTM. In future we probably want to change the visualization of confusion matrix to make it more understandable with that many classes.
|
Agreed Tolga. This PR just makes it no longer break the tool. But it could be improved with a more significant confusion matrix change. @stephanwlee please review |
| this.selected.length > 0 ? this.selected[0] : 0; | ||
| // Get the class labels of the top scoring classes from the most | ||
| // recent inference. | ||
| const labels = this.examplesAndInferences[exampleId].inferences[ |
There was a problem hiding this comment.
Woah, this is a bit hard to read and parse. How about something like
const currentExample = this.examplesAndInferences[exampleId];
const lastInference = currentExample.inferences[currentExample.inferences.length - 1];
// or
// const lastInference = currentExample.inferences.slice(-1)[0];
const labels = lastInference[modelIdx].slice(0, this.maxInferenceEntriesPerRun);| for (let modelIdx = 0; modelIdx < data.length; modelIdx++) { | ||
| let modelDataToChart = {}; | ||
| if (Object.keys(data[modelIdx]).length > | ||
| parseInt(this.maxInferenceEntriesPerRun)) { |
There was a problem hiding this comment.
parseInt should have a radix or you are supposed to use Number(this.maxInferenceEntriesPerRun) unless base is not 10.
There was a problem hiding this comment.
+1, but stronger: parseInt should have a radix, but even then Number
is preferred by Google style, because parseInt("1 2 3") is 1 while
Number("1 2 3") is NaN. (Our internal TS style guide says “must not”
use parseInt.)
5 participants
Models with many classes (like 100) have display issues in WIT
Only show top N classes in PD plots, just like we already do in inference panel
Correctly scroll large confusion matrices instead of overflowing into area where it can't be viewed
Ran all demos to see no regressions
Ran a demo with 100 classes to verify better visuals